Artificial intelligence-driven precision treatment of reproductive medicine-related diseases: the optimal protocol choice for IVF-ET

Graphical abstract

An AI-driven clinical decision support system (CDSS) personalizes ovarian stimulation (OS) for IVF-ET. Inputs: baseline demographics, infertility etiology, day-3 labs, and ultrasound. Engine: a data-driven, adaptive ensemble (ACA-FI for feature evaluation, and IRF for prediction) simulates hCG-day P, E2, EMT, and NOR under candidate OS protocols (antagonist, long, ultra-long agonist). Evaluation: predictions are mapped to a pregnancy grading scale (Levels I–IV) and integrated with time–cost metrics. Output: the CDSS recommends the highest-value OS, displaying predicted outcomes in a clinician-friendly dashboard. Impact: enables higher pregnancy probability with lower cost and fewer treatment days.

Highlights

• •AI-driven clinical decision support system (CDSS) for IVF-ET: efficacy–economy–time personalized OS selection.
• •Methodological features: data-driven, adaptive ensemble; robust, extensible.
• •The CDSS boosts clinical pregnancy rates while lowering cycles’ costs.
• •Identifies GnRH antagonist–suitable patients.

Abstract

Introduction

Optimal ovarian stimulation (OS) selection is critical for IVF success, but expert-based decisions often lack consistency in outcomes, cost-efficiency, and personalization, highlighting the need for more individualized and data-driven approaches.

Objectives

This study propose an artificial intelligence (AI) system that analyzes extensive IVF-ET cycles to uncover OS-pregnancy outcome relationships, enabling personalized treatment recommendations while improving success rates and minimizing unnecessary costs.

Methods

This study analyzed anonymized data from 17,791 patients undergoing OS and IVF/ICSI at Tongji Hospital between May 2015 and May 2019. An adaptive AI model was developed to predict key indicators—including progesterone (P), number of oocytes retrieved (NOR), estradiol (E2), and endometrial thickness (EMT) on the hCG day—by integrating personal characteristics, ovarian reserve, and etiological factors. This model facilitated personalized OS selection, pregnancy outcome grading, and the development of an AI-driven clinical decision support system (CDSS).

Results

The key indicators—progesterone (P), number of oocytes retrieved (NOR), estradiol (E2), and endometrial thickness (EMT) on the hCG day—were used to establish a pregnancy grading system. Pregnancy rates are stratified as follows: Level IV (Total Score 15–16), 0.55; Level III (Total Score 13–14), 0.44; Level II (Total Score 11–12), 0.24; and Level I (Total Score 4–10), 0.07. After OS optimization, 1,355 patients who were initially at level I were elevated to a better level. Of the 2,341 patients initially in level II, 2,290 improved, and of the 3,839 initially in level III, 1,448 improved. Patients elevated to level IV accounted for 80 percent of all cases. The CDSS prioritized a GnRH antagonist regimen for 54.64 % of patients, resulting in per-patient time savings of 15.39–33.48 days and cost reductions of ¥989–¥2,623 compared to non-optimal to antagonist. Scaled to China’s > 1 million ART cycles annually, this corresponds to projected direct savings of approximately ¥0.54–1.43 billion per year. In the new evaluation datasets (n = 4,251), implementation of CDSS recommendations increased the clinical pregnancy rate from 0.452 to 0.512 (p < 0.001) and reduced mean per-cycle cost from ¥7,385 to ¥7,242 (p = 0.018), demonstarting cost-effectiveness dominance with ICER saving of ¥2,383 per additional clinical pregnancy.

Conclusion

This AI-assisted CDSS streamlines clinicians’ decision-making by enabling efficient and accurate initial judgments on OS, standardizing and personalizing recommendations, and optimizing OS for effectiveness and cost-efficiency.

Introduction

In vitro fertilization and embryo transfer (IVF-ET), a primary assisted reproductive technology (ART), have been continuously advancing in addressing infertility issues. Since 1978, the International Committee Monitoring Assisted Reproductive Technologies estimated that at least 12 million babies have been born through IVF-ET and other ART procedures. In 2019, the success rate of achieving a live birth from over 3.41 million ART cycles reported by 81 countries globally was 22 %, with nearly one-third of the cycles coming from China [1]. However, the success rate of major ART procedures, particularly IVF-ET live births, remains relatively low. The process of IVF-ET is complex, involving multiple stages such as ovarian stimulation, egg retrieval, and embryo transfer. Each stage requires the experience and expertise of clinical professionals to make informed decisions. Additionally, the high cost of in vitro fertilization and embryo transfer poses a significant barrier, with limited medical insurance coverage in most countries, especially in low- and middle-income nations [2]. For instance, in China, the average cost of IVF-ET ranges from $8,108 to $300,000 up to May 21, 2024(China's implementation of assisted reproductive technology (ART) coverage under medical insurance policy on that day), all of which must be borne by the patients themselves.

Among the multiple stages of IVF-ET—including ovarian stimulation (OS), oocyte retrieval, fertilization, embryo culture, and embryo transfer—OS represents the earliest and most modifiable clinical intervention that directly influences downstream outcomes [3] Unlike later stages that rely more on biological variability and laboratory conditions, OS is predominantly protocol-driven and can be precisely tailored based on patient characteristics [4]. Several studies have demonstrated that the selection and timing of OS protocols significantly affect oocyte yield, embryo quality, endometrial receptivity, and ultimately, pregnancy success rates [5,6]. Moreover, OS is the single largest contributor to the overall cost and time of an IVF cycle. It accounts for approximately 30–40 % of direct treatment costs between $4,000 and $7,000 and determines the duration of treatment exposure to gonadotropins[3]. Inappropriate or suboptimal OS choices can lead to poor ovarian response, excessive stimulation, cycle cancellation, or complications such as ovarian hyperstimulation syndrome (OHSS), all of which compromise patient safety and economic efficiency [7]. The GnRH agonists long protocols are popular in China, because they have a lot of benefits, such as enhancing endometrial receptivity and increasing pregnancy rate of fresh IVF-ET cycles [8]. Compared with the GnRH agonists protocols, the GnRH antagonists protocols could offer shorter durations of analog treatment, shorter stimulation with FSH, less expensive, and a lower risk for ovarian hyperstimulation syndrome (OHSS). The ESHER (European Society of Human Reproduction and Embryology) recommended antagonist protocol for women with PCOS or predicted normal responder with regards to improved safety and equal eicacy [9]. But a meta-analysis showed that ongoing pregnancy rate was significantly lower in the antagonist protocols group compared with the agonist protocols group [9]. The current evidence suggested that these two medications were equally recommended for predicted poor responders [10]. Since 2014, the antagonist protocol has been increasingly favored and attracted attention from various reproductive centers in China due to its characteristics of simplified treatment regimens, reduced medication dosages, shortened treatment durations, and safety and efficacy. However, whether the antagonist protocol is the optimal choice for the majority of infertility patients remains uncertain.While previous Artificial intelligence (AI)-assisted research has focused largely on embryo selection or sperm classification, relatively fewer studies have addressed protocol-level decision-making at the OS stage [11]. Given that OS protocol selection is often based on empirical rules or physician experience rather than standardized algorithms, it represents a high-impact yet under-optimized decision node [12]. It is highly desirable to develop a tailored OS protocol and establish pregnancy evaluation under different OS scenarios to provide the best health care to each patient who needs IVF-ET.

Biomedical data collected from patients, laboratory tests, imaging examinations, and other sources has increased dramatically [13,14]. However, it is difficult for clinicians to extract all the meaningful data from the explosion of available biomedical data using conventional statistical methods [[15], [16], [17], [18], [19]]. AI technology and machine learning algorithms offer advantages in analyzing massive biomedical data and enhancing the understanding of disease diagnosis, visual recognition, and treatment outcome prediction [[20], [21], [22], [23], [24], [25]]. In IVF-ET, several studies have used deep neural networks (DNNs) to evaluate the quality of embryos and oocytes, and support vector machine (SVM) to select sperm cells, ensuring definition of good noninvasive markers and ART prediction models [[26], [27], [28], [29]]. However, the current research focus on development of an interpretable machine learning model for individualized gonadotrophin starting dose selection and the optimal day of trigger, but a reliable and detailed guidance and evaluation for Individual OS protocol is not available yet [[30], [31], [32]]. Furthermore, previous studies have applied a single approach to predict pregnancy outcomes and assist in reproductive medicine diagnosis and treatment, which can lead to biases and drawbacks.

This study designed a model system with an adaptive combination AI method to enable personalized prediction of pregnancy outcomes for each patient under different OS strategies and to propose the most cost-effective optimal OS cycles. The overall road map and outputs of the study are illustrated in Fig. 1. Firstly, an Adaptive Combined Algorithm for feature importance (ACA-FI) is expanded to provide importance ranking of features influencing IVF-ET results. Secondly, a pregnancy grading evaluation criterion is established based on the performance of the key indicators. Finally, an iterative random forest algorithm is utilized to predict the optimal OS selection and output the prediction of pregnancy outcomes. The predictive and evaluative system developed based on patient treatment and IVF-ET cycles provides evidence-based guidance on OS. The system can make the decision-making process for patients and clinical doctors more scientific and effective based on the clinical characteristics provided before OS treatment.

Fig. 1.

Schematic Diagram of Joint Design Approach for CDSS for OS Protocols and Pregnancy Outcome Prediction in Clinical Diagnosis and Treatment. The core clinical diagnosis and treatment process consists of five steps: I. personal information, II. ovarian reserve function and other etiology, III. OS, IV. follicle and hormone monitor on the HCG day, V. embryo monitor and pregnancy results. Firstly, an Adaptive Combined Algorithm for feature importance (ACA-FI) is expanded to provide importance ranking of features influencing IVF-ET results based on 13,540 patients at the Reproductive Medicine Centre of Tongji Hospital. Secondly, a pregnancy grading evaluation criterion is established based on the performance of the key indicators. Finally, an iterative random forest algorithm is utilized to predict the optimal OS selection and output the prediction of pregnancy outcomes and validated and optimized the model in 4,251 patients.

Methods

Description of the treatment procedure and data acquisition

Study design. A large cohort of anonymized register data collected from May 2015 to May 2023 at the Reproductive Medicine Centre of Tongji Hospital. The Centre ranks among the top ten reproductive medicine institutions in China, with its annual volume of IVF-ET cycles placing it within the leading cohort of reproductive centers nationwide. A retrospective cohort including 13,540 patients treated between May 2015 and May 2019 was used to develop and train the model. In addition, a prospective cohort comprising 4,251 patients treated between May 2019 and May 2023 was collected and used as an validation dataset. A total of 17,791 patients received OS and underwent IVF/ICSI therapy were included and analyzed.

Patients included in this study were selected from a prospective clinical database based on the following criteria: (1) Completion of one full OS cycle with embryo transfer; (2) Availability of baseline characteristics including age, BMI, AMH, Antral Follicle Counting (AFC), follicle-stimulating hormone (FSH), Inhibin B (INHB), cause of infertility, progestational hormone (P), estrogen (E2), endometrial thickness (EMT) and luteinizing hormone (LH).The hormonal assays (FSH, LH, E2, P, INHB) were conducted under International Organization for Standardization (ISO) 15,189 style quality [33], and ultrasound measurements for AFC/EMT was standardized with operator credentialing and periodic inter-observer calibration [34]. (3) Use of one of OS protocols to ensure treatment consistency. Patients with missing critical data or incomplete cycle outcomes were excluded. This approach ensures model training on high-quality, clinically meaningful cases with comparable treatment logic.

The study was conducted following the Standards for Reporting of Diagnostic Accuracy (STARD) reporting guidelines and approved by the ethics committee at Medicine Center. This prospective study conformed to the tenets of the Declaration of Helsinki.

Study data acquisition. An infertility diagnostic evaluation and ICSI/IVF treatment data analysis was conducted retrospectively, including personal information, ovarian reserve function and other etiology, OS, follicle and hormone monitor, and embryo monitor. For convenience, we name them process I, II, III, IV, V, respectively.

Ovarian stimulation (OS) protocols. The seven OS regimens are presented in Table S4 of the Supplementary Information as follows: long GnRH agonist protocol (M1), GnRH agonist short protocol (M2), antagonist protocol (M3), ultra-long GnRH agonist protocol (M4), the others (M5), mild OS with Letrozole (M6), Lutealphase ovarian stimulation protocol (M7). The clinical management of OS in this centre follows the recommendations of ESHRE (European Society of Human Reproduction and Embryology) [35]. The agonist stimulation protocol (M1) consisted of the administration of the GnRH agonist (Decapeptyl; Ferring, SaintPrex, Switzerland) (0.1 mg/day) from day 2 of the menstrual cycle followed by 150–225 IU/day recombinant FSH (Gonal-F; Merck-Serono, Geneva, Switzerland) from day 4 of the cycle [36,37]. GnRHa ultra-long protocol (M4) was performed by using ibidem triptorelin acetate 3.75 mg subcutaneously at cycle day 1–3 [38]. In the antagonist protocol (M3), the ovarian stimulation began with 225 IU of recombinant FSH from day 2 of the menstrual cycle, and the GnRH antagonist (Merck-Serono, Geneva, Switzerland) (0.25 mg/day) was introduced on the sixth day of the stimulation according to a multiple-dose, fixed protocol. Two leading follicles whose diameters reached 18 mm or three follicles whose diameters reached 17 mm were used to trigger ovulation with recombinant hCG (250 μg; Ovidrel; Merck-Serono, Geneva, Switzerland) [38]. Oocytes were retrieved transvaginally 34–36 h after hCG injection.

IVF procedures. Fertilization, embryo culture, and embryo transfer were performed according to the Vienna consensus/Istanbul updates for laboratory practices [39,40]. Fertilization was performed either by IVF or ICSI. Fertilization success was assessed 18h after insemination and the presence of two pronuclei (2PN) was regarded as normal fertilization. Fertilized zygotes were cultured until cleavage at day 3, followed by 2 or 3 days in G2-plus medium (Vitrolife, Sweden) to blastocysts. Day 2 and 3 Embryo development was assessed by counting blastomeres, assessing fragmentation, and examining blastomere symmetry. A rise in serum hCG levels on two consecutive occasions from 14 days after transfer indicated pregnancy. Each pregnancy with at least one intrauterine sac revealed by ultrasonography approximately 5 weeks after transfer was considered as a clinical pregnancy [37].

Data collection. Data were retrospectively collected from all patients during the process of IVF-ET. Fristly, personal information included age, BMI, duration of infertility, types of infertility, cycles of IVF. Secondly, ovarian reserve indicators and other etiology examination: AMH, Inhibin B (INHB), Follicle-Stimulating Hormone (FSH), Luteinizing Hormone (LH), estradiol (E2), progesterone (P), endometrial thickness (EMT), Antral follicle count (AFC), on the third day of menstruation. Thirdly, the aetiology is pointed by doctors through detailed clinical examination to tailor the ovulation stimulation protocol for the patient. The main etiologies of infertility includes A: Endocrine abnormality, B: Tubal factor, C: The uterus problem, D: The male problem, E: Chromosome abnormality/habitual abortion, F: Unknown cause. Fourthly, hCG was administered to ovulate, upon the development of a leading follicle and hormone monitor, including E2, P, Endometrial thickness, number of mature ovarian follicles. Fifthly, hCG treatment is followed by oocyte retrieval, in vitro fertilization, embryo monitor. The following evaluation indicators were used: number of oocytes retrieved, MII, 2PN, cleavage stage, Late cleavage stage, blastocyst yield. In depth analyses of associations of basic information with efficacy, the correlation among beta diversity metrics, this laies the foundation upon the pregnancy prediction and decision support model.

All the indicators used in the study and their explanations are presented in Table 1 and Table S4 (Supplementary file). The order of the processes represents the entire treatment cycle of IVF-ET, where processes I and II provide the patient’s basic information and ovarian reserve before treatment, process III involves the decision making for OS therapy, and processes IV and V contain follicle and embryo monitor. Abbreviations will be used to represent figures and text for convenience in text.

Table 1.

Characteristics of the dataset.

Description	Value(Mean ± sd)	Process
pregnancy results, n (%)	Pregnancy / Non-pregnancy (5149/8391)	Result
Pregnancy	5149 (38.03)
Non-pregnancy	8391 (61.97)
Cycles of IVF, n (%)		I
First	11189 (82.64)
Multiple	2351 (17.36)
Age, year	31.270 ± 4.974	I
Type of infertility, n (%)		I
Primary infertility	8422 (62.20)
Secondary infertility	5118 (37.80)
Duration of infertility, year	3.528 ± 2.678	I
Body Mass Index (BMI), kg/m2	21.882 ± 2.919	I
Baseline of Follicle-stimulating hormone (FSH), mIU/ml	7.769 ± 2.665	II
Antral follicle count (AFC), n (%)		II
AFC>=5	12430 (91.80)
AFC < 5	1110 (8.20)
Anti-Müllerian hormone (AMH), ng/ml	5.426 ± 4.186	II
Inhibin (INHB), pg/ml	91.608 ± 55.513	II
Luteinizing hormone (LH), mIU/ml	9.063 ± 149.041	II
Baseline of Estradiol (E2), pg/ml	63.582 ± 679.967	II
Baseline of Progesterone (P), ng/ml	0.939 ± 9.615	II
Baseline of Endometrial thickness (EMT), mm	5.645 ± 11.229	II
Causes of infertility, n(%)		III
Endocrine abnormality (A)	2131 (15.74 %)
Tubal factor (B)	3873 (28.60 %)
The uterus problem (C)	755 (5.58 %)
The male problem (D)	1963 (14.50 %)
Chromosome abnormality/habitual abortion (E)	868 (6.4 %)
Unknown or multiple cause (F)	3950 (29.17 %)
Ovarian stimulation protocols (OS),n (%)		III
Long protocol (M1)	3894 (28.62 %)
Short protocol (M2)	4 (0.02 %)
Antagonist protocol (M3)	3406 (25.16 %)
Ultra-long protocol (M4)	5183 (38.28 %)
The Others (M5)	222 (1.64 %)
Mild ovarian stimulation with Letrozole (M6)	822 (6.07 %)
Luteal-phase ovarian stimulation protocol (M7)	54 (0.398 %)
Number of mature ovarian follicles (NMOF), n (%)		IV
E2 at hCG day, pg/ml	3015.803 ± 1952.083	IV
P at hCG day, ng/ml	0.994 ± 0.940	IV
EMT at hCG day, mm	11.519 ± 2.820	IV
Number of oocytes retrieved (NOR), n	12.2 ± 6.794	V
Metaphase II (MII), n	10.701 ± 6.133	V
2 pronuclei (2PN), n	7.353 ± 4.613	V
Single pronucleus (1PN), n	0.445 ± 0.782	V
Multiple pronuclei (MPN), n	0.882 ± 1.427	V
Cleavage stage, n	8.516 ± 5.315	V
Late Cleavage, n	0.329 ± 0.809	V
Blastocyst yield, n	3.940 ± 3.987	V

Data Preprocessing. To ensure data integrity and model robustness, we performed standardized preprocessing steps prior to model training: (1) Missing Data Imputation: For continuous variables (e.g., age, BMI, FSH), missing values were imputed using the median of the non-missing values, which is robust to skewed distributions. For categorical variables (such as infertility type and postoperative follow-up regimen), samples with missing values were excluded during the initial data preprocessing stage. Variables with missing data exceeding 10 % of total observations were excluded from the modeling process to avoid bias from excessive imputation. (2) Outlier Detection and Removal: Outliers were defined as values falling beyond 3 standard deviations (SD) from the variable’s mean. For critical clinical parameters, we reviewed the source medical records. If values were confirmed as data entry errors or biologically implausible, they were excluded. These steps ensured both clinical plausibility and statistical consistency in the training and validation cohorts.

Multicollinearity Assessment. Here, we conduct a descriptive analysis to illustrate the presence of collinearity among features, providing a reference for subsequent model development. Specifically, Fig. 2A presents the correlation matrix of all continuous numerical features used in the study. Overall, during the pre-OS (Process I and II) phase, a relatively strong correlation is observed between AFC and AMH (Pearson correlation coefficient = 0.71), while correlations among other variables are generally weak. In addition, both AFC and AMH show relatively strong correlations with several post-OS variables, including NMOF, E2, NOR, MII, 2PN, and CLS. Among the post-OS variables, NMOF also exhibits strong correlations with E2 and NOR. Taken together, these findings suggest the presence of moderate collinearity among a limited subset of features in the dataset.

Fig. 2.

Descriptive analysis of IVF-ET process. A. A correlation matrix of the overall numeric features, where darker areas represent stronger correlations (blue is positive, purple is negative). B-C. The IVF success rate under different causes and OS. The larger the sector radius, the greater the success rate. D. The bubble chart of the factors importance in the whole IVF-ET process, and the size of the bubble represents the importance value, ranging from 0 to 1. E2(b), P(b), EMT(b) are on baseline Day 3 of menstruation, E2, P and EMT are on the day of human chorionic gonadotropin (hCG) administration. E. The distribution of age for IVF-ET cycles. F-I. The values of the top-four features (P, NOR, E2, EMT) in pregnant or non-pregnant groups and the distribution of pregnancy rates.

In addition, we further assess the correlations among features and their potential impact on the results using the Variance Inflation Factor (VIF). To assess multicollinearity among predictor variables, VIFs were computed by regressing each predictor variable against all others in turn. The VIF values quantify the degree of multicollinearity for each predictor. The detailed results are presented in Table S3 (Supplementary file). The results show that all predictor variables have variance inflation factors (VIFs) below 2, indicating low multicollinearity among the features and minimal impact on model stability.

Statistical analysis

The method of this study can be divided into three main steps. First, the ACA-FI method is used to assess the importance of features that affect the outcomes of in IVF-ET and select key features. Secondly, the key features are used to establish the efficiency of OS and pregnancy grading. Finally, a treatment pathway model for OS is learned using random forest (RF), and further iterative calculations are performed to obtain the optimal values for OS and related indicators.

Adaptive combined algorithm for feature importance (ACA-FI)

In this section, a combined method was used for evaluating feature importance, aiming to reduce the instability of single method feature screening [41]. We extend it to the case of binary response variables. All of these methods are capable of assessing feature importance and assigning a corresponding importance score. This process will thoroughly combine the methods’ advantages and make the feature importance ranking more reliable.

The calculation process. Assume that there are n independent observations $Y={\left(Y_1,\cdots ,Y_n\right)}^T$ and $X={\left(X_1,\cdots ,X_n\right)}^T$ be the feature vector and $X_i=(X_{i1},...,X_{\mathit{ip}})$, where $Y_i$ follows a Bernoulli distribution with conditional probabilities $P\left(Y=1|X=x\right)=f\left(x\right)$ given $X=x$, the regression function f need to be estimated by each method and its form is unknown. Referring to the process of the SOIL [42], we extend its calculation framework to an adaptive combined method feature importance (ACA-FI). Specifically, the procedure involves the following three steps:

Step 1. The feature importance scores $F{I}_k$ are computed for each p features under the $k$-th method, and each $F{I}_k$ has been standardized to output values between 0 and 1. By default, all methods are applied with their standard parameter settings.

Step 2. Calculate the corresponding weight $w_k$ for each recommended feature based on each method's prediction.

Step 3. Finally, A adaptive combined feature importance score with K methods is given by

$$\mathit{ACFI}={\sum }_{k=1}^K{w}_{k}F{I}_k.$$

In this paper, we take K = 3, and the calculation of weight of the kth method $w_k$ as follows. Specially, we combine three popular procedures: Random Forest (RF), XGBoost (XGB) and Sparsity Oriented Importance Learning (SOIL) [[42], [43], [44]]. These methods are widely used for feature importance evaluation. Among them, RF and XGB are non-parametric approaches that perform well on complex, nonlinear datasets, while SOIL is more suitable for parametric structures and interpretable relationships. The inclusion of these heterogeneous methods allows ACA-FI to handle diverse data characteristics and enhances its ability to generalize to new or unseen datasets. By leveraging the complementary strengths of these algorithms, the framework increases its robustness and adaptability in different clinical or population settings. Each with specific parameter settings as follows: For XGBoost classifier, a tree-based boosting model was trained using the caret package with 5-fold cross-validation. A grid search was conducted across multiple hyperparameters, including nrounds (1–100, step 20), eta (0.01–0.2, step 0.04), gamma (0.1–3, step 0.5), max_depth (1–20, step 4), with colsample_bytree, min_child_weight, and subsample fixed at 1, 1, and 0.5, respectively. For RF model, variable importance was assessed using the Gini impurity index, with the number of trees (ntree) set to 200 while all other parameters were kept at their default settings. For the SOIL method, default parameter configurations were applied throughout.

Weighting calculation. Assuming that each feature importance method assigns a standardized importance score to all features, variables with very low importance scores according to each method were first removed. Let the set of important features identified by method k be denoted as $X^k.$ Then, we estimate each model function based on these selected features, a list of model estimators $\widehat{f_k}\left(X\right)$ for f under each method with important features $X^k$. Next, we perform weight estimation using random data splitting to evaluate the predictive performance of each method. Methods that perform well in prediction will be assigned higher weights, while those with poorer performance will receive lower weights. This process is entirely dependent on the compatibility between the data and each method, making it an adaptive training approach. For t-th random data-splitting, $t=1,2,...,T$.

Step (1): Splitting the data into two parts, $D_1={\left(X_i^k,Y_i\right)}_{i=1}^{n/2}$ and $D_2={\left(X_i^k,Y_i\right)}_{i=n/2+1}^n$

Step (2): Obtain the estimators $\widehat{f_k}\left({X_i}^k\right)$ based on $D_1$.

Step (3): Assess the accuracy of the model using the remaining half of the data $D_2$. For each model, predict $Y_i$ by $\widehat{f_k}\left({X_i}^k\right),i=\frac{n}{2}+1,...,n.$

Step (4): Calculate the weight

$$w_k^{\left(t\right)}=\frac{{\prod }_{i=n/2+1}^n\widehat{f_k}{\left({X_i}^k\right)}^{Y_i}{\left(1-\widehat{f_k}\left({X_i}^k\right)\right)}^{\left(1-Y_i\right)}}{{\sum }_{l\in K}{\prod }_{i=n/2+1}^n\widehat{f_l}{\left({X_i}^k\right)}^{Y_i}{\left(1-\widehat{f_l}\left({X_i}^k\right)\right)}^{\left(1-Y_i\right)}}$$

Then, $w_k=\frac{1}{T}{\sum }_{t=1}^T{w}_k^{\left(t\right)}$ is the weighting estimation for $T$ data-splittings.

The ACA-FI framework is designed to be fully adaptive to data structure and clinical context. Because model weights are determined dynamically through performance-based averaging across repeated random data-splitting, the final weighting pattern reflects the intrinsic compatibility between each algorithm and the dataset, rather than any pre-specified preference. This ensures that the ensemble’s effectiveness does not rely on a specific data source or institution but instead captures the generalizable feature–outcome relationships embedded in the data. In other datasets or multi-center settings, the weighting process will automatically adjust according to each method’s predictive performance, allowing ACA-FI to maintain robustness and adaptability across heterogeneous populations [42]. This data-driven adaptivity also mitigates potential biases introduced by single-center data and supports the broader applicability of the framework to diverse clinical environments.

Here we first selected the features with importance large than 0.02 under the method $k$, and we set $T=100$ random sampling. The rationale for using the 0.02 threshold in this study was based on our observation that, in the multivariable context, a small subset of features consistently demonstrated very low importance scores across different methods. To reduce redundancy, we excluded these features at an early stage based on the recommendations of each individual method. For other datasets, the threshold can be flexibly determined by inspecting the distribution curve of feature importance scores and selecting an appropriate cutoff point accordingly. To determine the contribution of each model in the ensemble, we randomly split the data into two equal parts each time, with $n_1=n_2=\frac{n}{2}$.

In the analysis presented in this study, we employ Random Forest, XGBoost (XGB), and SOIL to compute adaptive feature importance scores through ACA-FI. The resulting weights assigned to each method are 0.95 for Random Forest, and 0.025 each for XGB and SOIL. These weights reflect the specific performance of each method on the dataset used in this study. In other words, although it is unclear a priori which method is most suitable for this dataset, the results indicate that the weighting scheme favors the performance of the RF model. The same weighting strategy can be applied to other datasets, where the results are determined in a data-driven manner—that is, without the need to predefine or subjectively select the most appropriate method. Instead, repeated data partitioning allows the model to adaptively determine optimal weights. This approach enhances generalizability and ensures that the algorithm produces stable results across different datasets. Here, we retain the results from all methods without excluding any, in order to preserve methodological diversity and enhance robustness. Additionally, the number of methods K can be expanded to include more approaches, provided they meet the necessary computational requirements. Fig. 2D illustrates the importance scores of all features based on the ACF-FI method.

Definition of OS efficacy evaluation

The OS efficacy evaluation based on the precise prediction and graded rating of E2, P, EMT and NOR on the hCG day after OS, which are categorized into four grades based on the pregnancy rate (PR). They were rated Excellent, with a pregnancy rate of 40 %; A Good pregnancy rate of 30 % to 40 %; Fair trade with a pregnancy rate of 20 % to 30 %; Poorly educated with a pregnancy rate of less than 20 %. Fig. 2 F-I shows the relationship between the values of the four indicators and the pregnancy rate. The standard is in accordance with previous research, clinical guideline and the distribution of cases in this study [[45], [46], [47]]. We take the scores for the four levels, denoted by 4, 3, 2 and 1. The grading criteria and scores are given in Table 2. The final total score for each patient is the sum of the four feature scores. Four levels of pregnancy grading were derived: Level I (Total Score of 4 to 10; PR is 0.07, range from 0 to 0.11), Level II (Total Score of 11 to 12; PR is 0.24, range from 0.18 to 0.29), Level III (Total Score of 13 to 14; PR is 0.44, range from 0.38 to 0.49), Level IV (Total Score of 15 to 16; PR is 0.55, range from 0.53 to 0.58).

Table 2.

Grading and scoring of the four features.

Feature	Performance	Score	pregnancy rate (PR)	Range of clinical indicator values	Non-pregnancy case count	Pregnancy case count	Total number of cases
P (pg/ml)	Poor	1	PR ≤ 20 %	P > 1.5.	1299	146	1445
	Good	3	30 %<PR ≤ 40 %	P ≤ 0.4,1.2 < P ≤ 1.5.	1772	920	2692
	Excellent	4	PR > 40 %	0.4 < P ≤ 1.2.	5320	4083	9403
E2 (ng/ml)	Poor	1	PR ≤ 20 %	E2 ≤ 500,E2 > 6800.	884	87	971
	Fair	2	20 %<PR ≤ 30 %	500 < E2 ≤ 1000,5300 < E2 ≤ 6800.	1351	521	1872
	Good	3	30 %<PR ≤ 40 %	1000 < E2 ≤ 1350,3150 < E2 ≤ 5300.	2990	1914	4904
	Excellent	4	PR > 40 %	1350 < E2 ≤ 3150.	3166	2627	5793
EMT (mm)	Poor	1	PR ≤ 20 %	EMT ≤ 8,EMT > 22.	1236	207	1443
	Good	3	30 %<PR ≤ 40 %	8 < EMT ≤ 11,20 < EMT ≤ 22.	2984	1667	4651
	Excellent	4	PR > 40 %	11 < EMT ≤ 20.	4171	3275	7446
NOR (number)	Poor	1	PR ≤ 20 %	NOR ≤ 3,NOR > 22.	1865	239	2104
	Fair	2	20 %<PR ≤ 30 %	3 < NOR ≤ 7,20 < NOR ≤ 22.	2112	1013	3125
	Excellent	4	PR > 40 %	7 < NOR ≤ 20.	4414	3897	8311

To isolate the independent effect of CDSS-recommended OS protocol on pregnancy outcomes, we performed multivariate logistic regression. The dependent variable was clinical pregnancy (yes/no), and the key independent variable was whether the patient received a CDSS-optimized OS protocol. Covariates included age, BMI, AMH, AFC, baseline FSH, infertility type (primary vs secondary), and infertility etiology (categorical). Adjusted odds ratios (aORs) and 95 % confidence intervals (CIs) were reported. Statistical significance was defined as p < 0.05.

Taking the GnRH agonist short protocol as the reference group, the analysis showed that both the GnRH antagonist protocol (adjusted OR = 0.775, 95 % CI: 0.700–0.859, p < 0.001) and the ultra-long protocol (adjusted OR = 1.563, 95 % CI: 1.392–1.437, p < 0.001) remained statistically significant predictors of pregnancy outcome. This confirms that the choice of OS protocol is an important and independent determinant of treatment success, even after adjusting for confounders.

Iterative Random Forest (IRF)

An iterative random forest algorithm with default parameter settings was used to predict the selection of the optimal OS protocol. We assume that the feature generation process after OS is as followed by the function $g\left(X\right)$, and the algorithm as follows:

Step 1: Draw a bootstrap sample of size $0.9n$ from the train data, the remaining data is used as test.

Step 2: Grow a random forest tree $\widehat{g_m}$ to the bootstrapped data based on given OS, by recursively repeating the following steps for each terminal node of the tree.

a. Select m variables at random from the $p$ variables.

b. Pick the best variable/split-point among the $m$.

Step 3: Output the ensemble of trees ${\left\{{\widehat{g}}_m\left(x\right)\right\}}^M$.

Step 4: $\widehat{g}\left(x\right)=\frac{1}{M}{\sum }_{m=1}^M\widehat{g_m}\left(x\right)$.

Step 5: Determine if the predicted $\widehat{g}\left(x\right)$ falls within the excellent range. If not, assign a new OS and repeat steps 1–4 until $\widehat{g}\left(x\right)$ falls within the ideal range for that particular feature.

Specifically, in this paper, we select the eleven features (Age, BMI, FSH, AFC, AMH, INHB, Cause, P(b) E2(b), EMT(b) and LH) before OS $\left(p=11\right)$ to predict the four features (Here $\widehat{g}\left(x\right)$ is respectively set to represent P, NOR, E2, EMT) after OS, respectively. Each Random Forest in the IRF framework was constructed using the default settings of the randomForest package in R, which include 500 trees (ntree = 500) and a maximum depth determined by full tree growth (i.e., until all terminal nodes are pure or contain fewer than the minimum number of samples). The number of variables randomly sampled at each split (mtrty) was set to the square root of the number of predictors, also following the default. The four indicators are divided into different levels based on their clinical values and their relationship with pregnancy rate. Please refer to Table 1 for detailed parameter settings, and the algorithm workflow is illustrated in Fig. 3A.

Fig. 3.

Implementation and evaluation of the Clinical Decision Support System (CDSS). A. The flow chart of OS selection algorithm and the clinical decision support system of ovarian. B-E. The change of the four features before and the optimal OS prediction. F. The grade distribution of patients before and after optimal OS. G-H. The optimal OS recommendation and the corresponding PR, cost, and time.

We design iterative algorithms for selecting the optimal OS protocol. Considering that the majority of patients use long GnRH agonist protocol, antagonist protocol, ultra-long GnRH agonist protocol, we selected 12,490 data points for training the OS optimization. The entire calculation is divided into two parts: optimization of OS protocols and calculate pregnancy score under optimal OS. First, we train the predictions of four indicators under different OS protocol choices using 11 features from the first two processes (Process 1 and Process 2) and using IRF to select OS protocols until the values of the four indicators reach the optimal range (The ideal condition is for each of the four indicators to fall within the excellent range. However, if they cannot be achieved, the option is to output results in the good range).

The convergence criterion for our algorithm was explicitly based on clinically established grading standards (as summarized in Table 2). Specifically, the algorithm computed the predicted outcomes for P, E2, EMT, and NOR for all considered OS protocols, and subsequently compared each protocol against these clinical thresholds. The algorithm selected the OS protocol achieving the highest grading across all four clinical indicators as the optimal solution. Once this best-performing protocol was identified, the iteration process was terminated, thus eliminating any risk of infinite loops or premature termination. Refer to Fig. 3A for the flowchart of the entire algorithm.

To evaluate the prediction accuracy of the IRF model, we calculated the mean squared prediction error (MSPE) for each of the four target outcomes: P, NOR, E2, and EMT (Table S1). Specifically, the following formula was applied:

$$\text{MSPE=mean}\left[{\left(\widehat{Y}-Y\right)}^2\right]$$

where $\widehat{Y}$ denotes the predicted value (RF output) and $Y$ denotes the observed true value. The logarithmic transformation was used to mitigate the influence of skewed distributions and large outliers, and to better capture relative prediction errors rather than absolute differences. This MSPE was computed on the test set across all target variables. The associated standard deviation (SD) of the prediction errors was also reported to reflect variability in model performance. This approach provides a more robust evaluation of prediction quality across different scales and magnitudes, and helps to ensure the stability and generalizability of the IRF model.

To evaluate the predictive performance of different machine learning algorithms, we implemented and compared two models: IRF and Gradient Boosting Machine (GBM). The dataset was randomly split into training and testing sets. Model hyperparameters were optimized using cross-validation on the training set. Model performance was assessed on the test set using three metrics: root mean squared error (RMSE), mean absolute error (MAE). Each model was trained and evaluated over multiple runs to account for variability, and the mean ± standard deviation of each metric was reported (Table S2).

Assessment of cost and time efficiency. After identifying efficacy‐optimal options based on the predicted pregnancy outcomes, the CDSS framework integrated two cost metrics to support comprehensive OS protocol optimization: Direct treatment cost and protocol duration.

• 1)Cost measurement: The cost for each OS cycle was calculated using actual patient billing records obtained from the financial department of Tongji Hospital. It included the full expense of gonadotropins, antagonist/agonist agents, monitoring visits, and laboratory procedures from stimulation initiation to embryo transfer. The average cost was computed separately for each protocol (M1, M3, M4), and values were used as reference benchmarks in the model’s output display.
• 2)Time measurement: Time duration was defined as the number of calendar days from the first day of ovarian stimulation to the day of embryo transfer, based on electronic scheduling logs. For example, antagonist protocols typically spanned 8–10 days, while long agonist cycles ranged between 20–25 days.
• 3)Value-based recommendation rule: When the IRF-based efficacy evaluation yields multiple protocol options with comparable predicted pregnancy outcomes, the CDSS applies a time–cost ranking to prioritize the recommendation: antagonist, long agonist, ultra-long agonist (see Section 3.4 for results). The final protocol is determined by the physician in discussion with the patient, refering to the CDSS recommendation.
• 4)ICER definition for economic comparison: We performed a within-patient economic evaluation comparing the original physician-recommended protocol versus the CDSS-optimized protocol for the same cohort (n = 4,251). Let $\mathrm{\Delta }C={\overline{C}}_{\mathit{CDSS}}-{\overline{C}}_{\mathit{Phys}}$ denoted the incremental direct cost per cycle, and $\mathrm{\Delta }E={\overline{p}}_{\mathit{CDSS}}-{\overline{p}}_{\mathit{Phys}}$ the incremental clinical pregnancy probability per cycle. The incremental cost-effectiveness ratio (ICER) is $\mathit{ICER}=\frac{\mathrm{\Delta }C}{\mathrm{\Delta }E}$ (incremental direct cost per additional clinical pregnancy per cycle) [48,49].

Calculation Instructions. In the process of data analysis, R version 3.6.2 (The R Project for statistical Computing, https://www.r-project.org/) was used to calculate all results covered in this paper, including descriptive analysis, feature importance and OS prediction. The R packages randomForest (https://www.stat.berkeley.edu/breiman/RandomForests/), xgboost (https://github.com/dmlc/xgboost), and SOIL (https://github.com/emeryyi/SOIL) are used to calculate feature importance for binary outcomes, and randomForest are also used to make predictions

The calculations for all three methods are performed using default settings. In terms of significance, P < 0.01 indicates high statistical significance.

Results

Clinical characteristics of a large cohort of IVF-ET patients

The cohort was 13,540 for part I (retrospective cohort) and 4,251 for part II (prospective cohort), with treatment records from Tongji Hospital, a renowned medical institution. The part I was training data for model building, while the part II was test dataset for model evaluation. The data in this study reflected the real-world case mix seen in reproductive medicine centers. Our training cohort comprised 13,450 cycles and was broad in scope and composition. It included 11,190 first cycles (82.64 %) and 2,351 repeat cycles (17.36 %); 8,422 primary infertility (62.20 %) and 5,118 secondary infertility (37.80 %). Etiologies were diverse: tubal factor (28.60 %), endocrine abnormalities (15.74 %), male factor (14.50 %), uterine factor (5.58 %), chromosomal abnormalities/recurrent miscarriage (6.40 %), and unknown or multiple causes (29.17 %). All major OS regimens in current practice were represented, including ultra-long GnRH agonist (38.28 %), long GnRH agonist (28.62 %), and GnRH antagonist (25.16 %) protocols, with the remainder receiving other regimens. These characteristics supported the cohort’s representativeness for contemporary reproductive medicine practice. Detailed clinical information was collected involved personal information (I), ovarian reserve function and other etiology (II), OS (III), follicle and hormone monitor (IV), and embryo monitor (V) (Fig. 1). First, 13,540 IVF-ET cycles as the training set, we calculated the correlation of the numerical features involved in the IVF-ET process (Fig. 2A). Clearly, the features involved in process I are all negatively correlated with processes II, IV, and V. More specifically, age, IVF-ET cycle, and FSH on the third day of menstruation were negatively correlated with ovarian reserve indicators and embryo quantity and quality. Conversely, AMH and AFC are positively correlated with the number and quality of embryos. The mean pregnancy rate per couple was 38.03 %, with 82.64 % of initial cycles of IVF-ET, 62.2 % primary infertility and 37.8 % secondary infertility in Table 1. Among the cases, ages were approximately normally distributed with a mean age of 31.27 ± 4.95 years, and range from 20 to 51 years old; however, the percentage pregnancy rate tended to decrease steadily with advancing age (Fig. 2E). The results show that 32 and 37 are two turning points: from the 32-year point, the pregnancy rate drops below 40 percent. At 37, the pregnancy rate drops below 30 percent. The pregnancy rate experienced the largest declines at the age of 32 and 37 years (5 % and 4 %) for women. Across the board, IVF-ET brings opportunities for infertility treatment, but outcomes still need to be improved, especially for the elderly.

Analysis of the features that affect the outcomes of IVF-ET

An adaptive combination algorithm (ACA-FI) was used to calculate the importance of 27 features in the IVF-ET process (Fig. 2D). The results indicate that progesterone (P) on hCG day exhibited the most significant effect on pregnancy outcomes among all the features, followed by NOR, E2, EMT on hCG day, and age. The importance of each of these attributes is greater than 0.5. In addition, AMH, E2(b), BMI, LH, OS, and FSH all have importance values around 0.5, and their influence on IVF-ET outcomes is also evident. It is worth noting that among the lowerranked features, the type and cause of infertility, and the cycle of IVF-ET had little effect on the outcomes, except for indicators related to the embryo monitoring phase. Further, similar to previous recognition, the effect of AMH and age on IVF-ET were significant [50,51]. Although AFC and AMH showed moderate correlation (Pearson r = 0.67), both were retained due to their unique predictive contributions when combined with age and FSH. Model performance remained stable, and no instability was observed in the presence of these correlated variables.

However, the effect of infertility cause on outcomes is slight, and pregnancy rates are comparable across multiple causes (Fig. 2B,C). Patients with different infertility causes and similar outcomes need to undergo testing for the effectiveness of the OS process. The mean importance scores as 0.276, 0.463, 0.322, 0.658, and 0.276 for the five treatment processes. The importance of five processes increased first and then decreased, and the follicle and hormone monitor (Process IV) has the most significant influence on IVF-ET outcomes. In addition, the effect of ovarian reserve function (Process II) on the final outcomes is obvious.

The grading criteria for IVF-ET outcomes and the clinical decision support system

The relationship between the values of the four most important indicators after OS and the pregnancy rates generally follows an irregular inverted U-shaped pattern (Fig. 2F, G, H, I). This significant sensitivity contributes to the establishment, grading, and results of the evaluation system (Table 2). Among them, for P (number of people with a score of 1 is 1,445, score of 3 is 2,692, score of 4 is 9,403), for E2 (number of people with a score of 1 is 971, score of 2 is 1,872, score of 3 is 4,904, score of 4 is 5,793), for EMT (number of people with a score of 1 is 1,443, score of 3 is 4,651, score of 4 is 7,446), and for NOR (number of people with a score of 1 is 2,104, score of 2 is 3,125, score of 4 is 8,311). The majority of individuals scored 4, accounting for approximately 69.4 %, 42.8 %, 54.9 %, and 61.4 % for each indicator, respectively. Based on this, the grading of pregnancy outcomes constructed based on the four indicators is reasonable. Table 3 demonstrates significant variations in pregnancy rates across different pregnancy grades. Level I (total score of 4 to 10, with 2,090 individuals and the average pregnancy rate of 0.07), Level 2 (total score of 11 to 12, with 2,679 individuals and the average pregnancy rate of 0.24), Level 3 (total score of 13 to 14, with 4,200 individuals and the average pregnancy rate of 0.44), and Level 4 (total score of 15 to 16, with 4,571 individuals and the average pregnancy rate of 0.55).

Table 3.

Pregnancy grading with the four key indicators.

Grading	Total Score	Non-pregnancy case count	Pregnancy case count	Total number of cases	Actual pregnancy rate range
I	[5,10]	1953	137	2090	0–0.11 (0.07)
II	[11,12]	2026	653	2679	0.18–0.29 (0.24)
III	[13,14]	2356	1844	4200	0.38–0.49 (0.44)
IV	[15,16]	2056	2515	4571	0.53–0.58 (0.55)

Thereafter, a clinical decision support system (CDSS) was established based on the results and tested in clinical trials in outpatient practice (Fig. 3A, Fig. 4). To clearly illustrate the originality and practical application of our approach, we developed an AI-Clinical Decision Framework. This framework encompasses several key steps: (1) inputting patient personal information, infertility causes, laboratory reports, and day 3 menstrual data; (2) using the IRF model to calculate the predicted HCG day E2, P, EMT, and NOR for different OS protocols; (3) iteratively evaluating these predictions until all four indicators meet the optimal clinical criteria, at which point the system outputs the recommended OS protocol along with the predicted pregnancy outcome.

Fig. 4.

The clinical decision support system user interface of ovarian stimulation protocol.

To facilitate clinical translation, a user-friendly CDSS interface was designed (Fig. 4). In this system, clinicians or patients input the relevant data, and the CDSS automatically calculates and displays the expected E2, P, EMT, and NOR values, as well as the predicted pregnancy outcome for each candidate protocol. The interface provides visualized comparisons between different protocol options and their respective outcomes, empowering clinicians to integrate AI recommendations with their own clinical expertise for personalized OS decision-making. Additionally, the interface allows patients to view their individualized prognosis in a transparent manner, supporting shared decision-making and realistic expectations regarding treatment results.

Validation of the clinical decision support system(CDSS) for pregnancy evaluation, live birh outcomes and optimal OS protocol

To better illustrate the originality and translational logic of our approach, we propose a structured AI-Clinical Decision Framework (see Fig. 4). This framework outlines the end-to-end flow of information from baseline patient data acquisition to algorithmic recommendation and clinical validation. The model was validated with 13,540 patients (retrospective cohort) using the following procedure.

Firstly, by iteratively computing the IRF, we adjust the OS to ensure that P, NOR, E2, and EMT post-OS fall within the better range defined in Table 2. The distribution of the original level ranges for the four metrics and the predicted level ranges after optimizing the OS selection are shown in Fig. 3B-E. It can be observed that a significant number of patients do not fall within the optimal range for these four metrics after the initial OS treatment. Specifically, approximately 30 % of patients with NOR and 18 % of patients with E2 were in the fair range, while approximately 10 % of patients with P and NOR were in the poor range. However, after OS optimization selection by CDSS, the predicted values for all four metrics fall almost within the good and excellent range. This suggests that while doctors and specialists are providing patients with better OS with current treatments, the analysis results suggest that they are not yet at their optimum. We determine the best OS selection based on the given baseline information of the patients. In addition, we calculated the distribution of pregnancy results scores for patients before and after the optimal OS. The results are shown in Fig. 3F under the level division specified in Table 2. Overall, after OS optimization, more patients had their pregnancy level raised to Level III or IV. Higher levels indicate a higher probability of pregnancy. Specifically, after OS optimization, 1355 patients who were initially at level I were elevated to a better level. Of the 2,341 patients initially in II, 2,290 improved, and of the 3,839 initially in III, 1,448 improved. Patients elevated to level IV accounted for 80 percent of all patients. We recommend the OS that enables all four features to achieve excellence as the optimal choice. The optimal recommendation selection of OS for all patients is summarized in detail in (Fig. 3G). Of these, 2,922 selected antagonist OS simultaneously, while 535, 3934 and 17 recommended long GnRH agonist protocol, antagonist protocol, ultra-long GnRH agonist protocol respectively. Additionally, to validate the clinical utility of the pregnancy grading system beyond initial pregnancy prediction, a sub-cohort of 1,438 patients was analyzed for whom live birth outcome data were available. Spearman test for trend confirmed a statistically significant association (p < 0.001), indicating that the grading system not only predicts clinical pregnancy but is also a meaningful indicator of live birth success.

Secondly, based on the scores for pregnancy prediction, economic and time cost were considered. As shown in Fig. 3H, per-protocol references indicated that the GnRH antagonist regimen is the most economical and time-efficient (¥6,396; 9.52 days) compared with long agonist (¥7,385; 24.91 days) and ultra-long agonist (¥10,008; 43 days). However, optimizing for clinical pregnancy does not imply a uniform antagonist recommendation (Fig. 3G). When multiple protocols achieved comparable predicted pregnancy outcomes, we rank CDSS outputs by value—antagonist first, then long agonist, and ultra-long agonist last. Under this rule, 54.64 % (6824/12490) of patients are prioritized for the antagonist regimen. For the antagonist-suitable subgroup, individualized recommendations yield 15.39–33.48 fewer stimulation days and ¥989–¥2,623 lower direct costs per patient relative to non-optimized antagonist, while maintaining or improving the predicted probability of pregnancy. Applied conservatively to the antagonist-suitable share (54.64 %) of > 1,000,000 ART cycles annually in China [52], this corresponds to projected direct savings of approximately ¥0.54–1.43 billion per year.

Thirdly,the performances of 10-fold cross-validation under 100 times on the IRF model are presented in Table S1. Table S1 shows the average prediction error for the four key metrics on the train and test datasets. The mean squared error results demonstrate that our predictions for the four metrics are highly accurate after using RF. Moreover, the prediction results are stable, ensuring the reliability of the OS prediction.

The performance metrics for the IRF and GBM models are summarized in Table S2. The table summarized the root mean squared error (RMSE), mean absolute error (MAE) of four target variables (P, NOR, E2, and EMT) for both the IRF and Gradient Boosting Machine (GBM) models. IRF model outperformed the GBM model in terms of both error metrics and explained variance for the four target variable.

Testing the clinical decision support system (CDSS) for optimal OS protocol

We collected a prospective queue comprising treatment information from 4251 patients from May 2019 to May 2023 at Tongji Hospital. Here, we have re-recommended individual OS for patients through our algorithm to improve their outcome grading while reducing costs. The specific comparison results are shown in Table 4. On this new dataset, based on the grading, we estimated that for patients who originally received the GnRH agonist protocol or antagonist treatment, their average success rates and costs were pregnancy rate = 0.46, Cost=¥7385 and pregnancy rate = 0.452, Cost=¥6396, respectively. However, with CDSS’s individual recommendations, out of the 417 patients who originally received only GnRH agonist protocol, we recommended GnRH agonist protocol for 55 patients, antagonist for 271 patients, and ultra-long GnRH agonist protocol for 88 patients. In this scenario, we calculated their pregnancy rate = 0.534, Cost=¥7242.67. Compared to solely using GnRH agonist protocol, not only did the success rate increase by 0.074, but the cost also decreased by ¥143. Similarly, out of the 3834 patients originally recommended antagonist, we allocated GnRH agonist protocol to 1074 patients, antagonist to 2257 patients, and ultra-long GnRH agonist protocol to 487 patients. Under the new circumstances, this group of patients achieved a pregnancy rate = 0.512, Cost=¥7089.49, with a 0.06 increase in success rate, despite recommending some more expensive ultra-long GnRH agonist treatments. To assess the significance of the observed improvements, we performed paired t-tests comparing the original physician-recommended protocols versus CDSS-optimized protocols on the same cohort of 4,251 patients. The increase in clinical pregnancy rate from 0.452 to 0.512 was statistically significant (p < 0.001). The reduction in average cost per cycle from ¥7,385 to ¥7,242 was also statistically significant (p = 0.018). The absolute pregnancy gain was + 6.0 percentage-point (relative + 13.3 %), and the average per-cycle cost decreased by ¥143 (1.9 %). On the cost-effectiveness plane, CDSS was dominant (higher effectiveness at lower cost), equivalently, the ICER indicates a cost saving of ¥2,383 for each additional clinical pregnancy.These findings support that even modest cost reductions, when applied across a large patient population, can yield meaningful economic benefits. Through our validation, it is evident that while antagonist may be a more common choice for OS, personalized recommendations yield higher success rates and lower costs, which is crucial.

Table 4.

Comparison of recommendations original and Induvidual on a new dataset.

		Original (PR = 0.46, Cost=¥7385)	New (PR = 0.534, Cost=¥7242.67)
	Grading (PR)	M1 (417)	M1 (55)	M3 (271)	M4 (88)
M1(417)	I (0.07)	10	0	0	0
	II (0.24)	56	0	0	0
	III (0.44)	123	6	25	16
	IV (0.55)	228	49	246	72
		Original (PR = 0.452, Cost=¥6396)	New (PR = 0.512, Cost=¥7089.49)
		M3 (3834)	M1 (1074)	M3 (2257)	M4 (487)
M3 (3834)	I (0.07)	183	0	0	0
	II (0.24)	505	14	17	1
	III (0.44)	1182	396	654	118
	IV (0.55)	1964	664	1586	368

Discussions

Despite IVF-ET being a significant breakthrough in treating infertility, pregnancy rate and delivery rate remain low. The low success rate of IVF-ET causes patients' mental and physical damage, economic and time losses. Additionally, the current guidelines do not delineate a recommended regimen for individual IVF-ET cycles. The selection of OS protocols largely relies on the experience of physicians. And discrepancies in OS protocols selection arising from differences in the levels of healthcare facilities and inexperienced practitioners across various reproductive medicine centers may impact the success rates of ART pregnancies. This study established a CDSS that could predict the IVF-ET results and assists the decision of customized optimal OS protocols. The CDSS could be recommended implemented in reproductive medicine centers to help doctors choose the best OS protocols, reduce young doctors' study cycle, increase the cost-effectiveness and pregnancy rates.

The CDSS is established based on large-scale data training and a rigorous scientific prediction process. We established the CDSS based on 13,540 patients at the Reproductive Medicine Centre of Tongji Hospital, and validated and optimized the model in 4,251 patients. First, we identified that E2, P, EMT, and NOR were the top four factors that were in relationship with pregnancy rate among the 27 factors (including personal information, ovarian reserve function and other etiology, OS, follicle and hormone and embryo monitor). The grading criteria for IVF-ET outcomes was built based on the relationship between four indicators and pregnancy rates. Most prior AI applications in ART have focused on laboratory-level prediction tasks, such as embryo viability assessment through time-lapse imaging, sperm selection, or oocyte grading using computer vision and deep learning techniques [53,54]. While these approaches optimize downstream steps of IVF-ET, they do not assist with pre-treatment protocol selection, which has a critical influence on both biological outcomes and patient burden. Our study is among the first to address this gap by introducing a CDSS that evaluates patient characteristics prior to stimulation and recommends personalized OS protocols, thereby shifting AI-driven decision support from the lab to the clinic. This upstream integration complements existing embryo-scoring models and represents a meaningful advancement toward fully individualized ART. Traditional OS selection relies heavily on clinician experience, qualitative interpretation of biomarkers, and institutional norms. While experienced physicians often achieve good outcomes, such approaches are inherently subjective and vary significantly across providers and centers. In contrast, our AI-driven CDSS systematically integrates multi-dimensional clinical data (e.g., AMH, AFC, age, FSH) and applies validated machine learning algorithms to generate consistent, individualized protocol recommendations. Furthermore, the CDSS also demonstrated lower inter-patient variance in cycle cost and gonadotropin dose. Combining with economic cost and time cost, the CDSS provides the maximum support to optimize the outcomes of OS, decreases the cost of trial and error, and improves their patients’adherence and pregnancy results.

The practical application of AI-based CDSS in reproductive medicine requires not only algorithmic accuracy but also seamless integration into real-world clinical workflows. Our CDSS was developed with this clinical operability in mind. The system is designed to interface with existing electronic medical records (EMRs) and outpatient registration systems, enabling automatic data extraction and processing of patients’ baseline information (e.g., age, AMH, AFC, hormone profiles) immediately after initial consultation. A user-friendly graphical interface has been developed to provide clinicians with real-time, data-driven OS protocol recommendations. These recommendations are based on the patient's clinical indicators and present a comparative assessment of protocol-specific predicted outcomes—such as E2, P, EMT, and NOR levels on the day of HCG—as well as estimated treatment costs, cycle duration, and efficiency. Clinicians can use this information alongside their own experience to evaluate and select the most appropriate protocol. These recommendations are generated in real time and are intended to serve as clinical references rather than prescriptive instructions. Importantly, the final treatment decision remains under the clinician’s authority, allowing for manual override in cases where clinical judgment deems it necessary due to nuances not captured by the algorithm (e.g., rare complications, psychological factors, or patient preferences). In addition, the system also provides predicted pregnancy probabilities, costs, and timelines to help patients form realistic expectations regarding their IVF treatment cycles. Furthermore, we identified key decision points in the workflow—such as protocol selection before stimulation initiation—where the CDSS outputs can be most effectively utilized. The implementation process also considers potential barriers, including clinician resistance to automation, data standardization across EMRs, and the need for physician training. To address these, pilot trials and clinician feedback loops were incorporated during the development phase. Overall, the CDSS is designed not to replace clinical expertise, but to augment it—especially in high-volume centers or settings with less experienced staff—by providing data-driven, evidence-based guidance that enhances decision-making efficiency and standardization.

We innovatively improved an adaptive combination method in the analysis of the relating factors, which can automatically combine different machine learning algorithms, and assign weights by how well fits the data for each method. The combination method fully integrates prediction accuracy of random forest, and the capacity to handle the missing data of XGBoost, while reduces the negative effect of flaw and bias of single method [43]. The integrated analysis of the indicators involved in the whole IVF-ET process of the case cohort provides important information. The ACA-FI ensemble framework is designed to be robust, data-driven, and extensible. Crucially, the framework is adaptive: weights are re-estimated when ported to new cohorts, so the ensemble can rebalance if other learners contribute more in different settings. This approach ensures the framework’s adaptability and relevance across different datasets and clinical contexts, reflecting its core purpose as a flexible tool for real-world protocol optimization. Our most research conclusion is consistent with previous conclusion, thereby validating the reliability of the model. Such as, the importance of AMH and age for IVF-ET, the positive correlation of AMH, AFC with embryo quantity and quality, and FSH, age, cycles of IVF-ET are negatively correlating with ovarian reserve indicators and embryo quantity and quality [50,55]. Based on these findings, this study optimizes the OS protocol through AI to improve the outcome of IVF-ET. However, in the current implementation, the methodological framework incorporated only three representative feature selection algorithms—RF, XGB, and SOIL—which may introduce a certain degree of selection bias. Model training and validation were conducted using data from a single large-scale center (n = 17,791), which may raise concerns regarding potential limitations in external validity and generalizability. In future work, we plan to extend the ACA-FI framework by incorporating a broader range of algorithms and validating its performance across multi-center and multi-population datasets to further evaluate its adaptability and robustness.

More interestingly, this CDSS provides us with new significant clinical insights. This study extensively investigates the enduring debate regarding the utilization of GnRH agonists and antagonists protocols. The ESHER recommended antagonist protocol for women with PCOS or predicted normal responders, while both equally for those with poor predicted prognosis [9,56]. According to a meta-analysis, the antagonist group had a significantly lower ongoing pregnancy rate than the agonist in IVF-ET cycles. The antagonist has been applied in China since 2014, due to its simplified treatment regimen, reduced medication usage, shortened treatment cycle, and safety and efficacy [[1], [52]]. However, current clinical guidelines provided limited guidance on precisely which patients should receive GnRH antagonist protocols. Against this backdrop, our CDSS predicted individualized pregnancy outcomes under multiple OS options and then recommended the protocol with the most favorable benefit–risk–cost profile for each patient. In our center, the system identified 54.64 % of patients priority to antagonists. Extrapolated to China’s exceed one million annual ART cycles [52], this suggests ¥0.54–1.43 billion in yearly savings, easing pressure healthcare services and public health spending. Because pricing varies across institutions and our estimates are derived from a single-center cohort, these figures should be viewed as approximate and interpreted with appropriate caution. These advantages are particularly valuable in low-resource settings and countries where IVF is not fully covered by insurance, thereby enhancing the affordability and accessibility of fertility care. Thus, the CDSS was designed based on the “efficacy-economic-time” comprehensive concept, so that treatment effects and cost are balanced. The “effectiveness − economy − time” CDSS system was applied to the new dataset (n = 4,251). In the test set (n = 4,251), comparing CDSS recommendations with prior physician-selected protocols yielded an ICER of cost saving of ¥2,383 per additional clinical pregnancy, indicating dominance—that is, higher effectiveness at lower cost. Scaled to service volume, the point estimates implyed approximately 60 additional clinical pregnancies and ¥143,000 saved per 1,000 cycles. Future work should extend to cost per live birth and cost-utility analyses.

The study facilitated personalized treatment and precision medicine. There are, however, certain limitations to this study. The primary limitation is the single-center design. Although the cohort is large (17,791 patients over eight years), spaned a broad range of infertility etiologies, and includes all major OS regimens, thereby enhancing representativeness. However, external database verification is still necessary. Thus,we are actively seeking collaborations with additional IVF centers across China and internationally to perform external validation, which will help evaluate the model's robustness and generalizability under different clinical conditions. Methodologically, ACA-FI learns is data-driven and adaptive: when applied to other centers, the weighting and calibration steps are re-estimated on datasets, enabling reducing the risk that single-center idiosyncrasies drive results. To further improve stability and generalizability, we will implement standardized quality controls across participating centers, including ISO 15189–style quality systems for hormone assays [33], ultrasound standardization with operator credentialing and inter-observer calibration, Vienna consensus/Istanbul updates for IVF-lab and embryo assessment [39,40], and ESHRE guidance to harmonize ovarian stimulation protocols [35]. These measures will be embedded prospectively in the planned multicenter validation. Future work will also include a federated learning framework to allow model training across institutions without compromising patient data privacy. Additionally, we acknowledge that not all patient-specific or contextual factors can be fully captured by the algorithm. Clinical decision-making in assisted reproduction is multifaceted and may involve considerations such as prior treatment response, endometrial receptivity history, patient preferences, emotional readiness, and logistical constraints. Our CDSS is therefore not intended to replace physician expertise, but rather to augment pre-treatment planning by presenting data-driven protocol options. It is not designed to suggest or enable in-cycle protocol changes, which remain under exclusive physician control and are rarely feasible in practice. Due to the complex procedure, we also need to optimize the other steps in IVF-ET more than OS. Further, we will focus on more social information and life effects to promote primary prevention of infertility. In conclusion, applications of CDSS will assist clinicians in optimizing OS therapy, predicting pregnancy, and improving IVF-ET outcomes. This will reduce the risk of repeat IVF-ET cycles and associated complications. In summary, the clinical application of this CDSS can help almost all reproductive medicine centers achieve top medical standards effectively, assist young doctors in rapidly gaining experience, and enable patients to receive the most cost-effective ovarian stimulation protocols.

Data and Code availability

The data that support the findings of this study are available on request from the corresponding author [Liao]. The data are not publicly available due to them containing privacy information that could compromise research participant privacy or consent, such as age, infertility, diseases history, etc.. The core codebase used to implement the ACA-FI algorithm and iterative random forest (IRF) prediction model has been made publicly available on GitHub: https://github.com/wendali1009/IVF-ET-prediction,

The repository includes R scripts for feature importance calculation, model training, cross-validation, and CDSS output simulation, along with documentation and usage instructions.

Ethics approval statement

The study was approved by the Ethics Committee of Tongji Hospital, Tongji Medical College of Huazhong University of Science and Technology. The procedures used in this study adhere to the tenets of the Declaration of Helsinki. According to the Institutional Review Board (IRB) of Tongji Hospital, our study was not subjected to ethics review, because all women participating in the study received routine IVF–ET treatment in the hospital and no additional intervention or sampling was performed.

No material from other sources was reproduced in this study.

This study is not a clinical trial and was not registered.

Funding statement

This research was funded by National Natural Science Foundation of China (NSFC) (Grant no, 82341120; 82372354; 72474218; 82501971).

CRediT author statement

Li Wen, Di Wu, Jinhan Ruan, Renjie Wang: Contributed to Conceptualization, Formal Analysis, and Writing − Original Draft, including the main data analysis, organization, and preparation of the manuscript. Rui Long, Rongkai Chen, Cheng Hu, Chenxiao Tian, Yichuan Zhang: Contributed to Data Curation, Visualization, and Investigation, including data preprocessing, figure preparation, and resource collection. Wei Pan, Lei Jin, Shujie Liao: Provided Supervision, Writing − Review & Editing, and overall Project Administration, guiding the study and critically revising the manuscript.

Declaration of competing interest

The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.

Appendix A. Supplementary data

The following are the Supplementary data to this article: