A method and system for establishing a model for predicting pregnancy rate in early onset ovarian insufficiency

Abstract

The present application relates to the technical field of pregnancy rate prediction, and particularly relates to a model establishment method and system for predicting pregnancy rate of early-onset ovarian dysfunction, which obtains clinical data of a patient, identifies four types of prediction key points, including basic physiological characteristics, ovarian reserve function, treatment intervention and pregnancy outcome correlation. When the total amount of key points exceeds a first threshold value, a dimensionality reduction improved feedforward neural network is constructed, and basic physiological characteristics and treatment intervention data are used as inputs to predict ovarian function and pregnancy outcome. When the number of key points is small, a graph neural network is used, a directed graph structure is constructed based on clustering and hierarchical relationship of each key point, and node state is dynamically updated based on basic physiological characteristics, so as to realize classification prediction of ovarian function and pregnancy risk, improve prediction accuracy and model interpretability, and provide personalized and intelligent decision support for fertility evaluation of POI patients.

Description

Technical Field

[0001] This invention relates to the field of pregnancy rate prediction technology, and in particular to a method and system for establishing a model to predict the pregnancy rate of early-onset ovarian insufficiency. Background Technology

[0002] Premature ovarian insufficiency (POI) refers to the decline in ovarian function in women before the age of 40, manifested as menstrual irregularities, elevated gonadotropin levels, and decreased estrogen levels, severely impacting fertility and reproductive health. With the accelerating pace of life, changing environmental factors, and delayed childbearing age, the incidence of POI is increasing year by year, becoming one of the important clinical problems affecting the fertility of women of childbearing age. For POI patients, assessing their chances of natural pregnancy or pregnancy after assisted reproductive technology (ART) is a crucial step in developing an individualized treatment plan. However, current clinical assessment methods mostly rely on single indicators (such as anti-Müllerian hormone (AMH), basal follicle-stimulating hormone (FSH), and antral follicle count (AFC), lacking a systematic integration of multidimensional factors and failing to comprehensively reflect the patient's fertility potential.

[0003] In recent years, artificial intelligence (AI) technology has been increasingly applied in medical prediction models, showing great potential, especially in the field of reproductive medicine. While traditional statistical models such as Logistic Regression and Cox Proportional Hazard Models can be used to predict pregnancy outcomes, they struggle to handle high-dimensional, nonlinear, and complex clinical data. Deep learning models, such as Feedforward Neural Networks (FNNs) and Graph Neural Networks (GNNs), can uncover deeper relationships between variables, improving prediction accuracy. However, existing models are mostly focused on the general infertile population, and dedicated prediction models for the specific POI (Potentially Infertile Individual) group are still scarce. Furthermore, existing models generally suffer from feature redundancy, poor model interpretability, and weak adaptability to small samples.

[0004] The above content is only used to help understand the technical solution of the present invention and does not represent an admission that the above content is prior art. Summary of the Invention

[0005] The main objective of this invention is to provide a method and system for establishing a model to predict the pregnancy rate of patients with early-onset ovarian insufficiency. This aims to solve the technical problem that existing technologies lack accurate modeling methods for the dynamic association of multiple factors in patients with early-onset ovarian insufficiency, making it difficult to achieve efficient, interpretable, and highly adaptable individualized prediction of pregnancy rates.

[0006] To achieve the above objectives, the present invention provides a method for establishing a model to predict the pregnancy rate of women with early-onset ovarian insufficiency, the method comprising: Acquire clinical data of patients with early-onset ovarian insufficiency, and determine the number of predictive key points and the corresponding predictive attribute categories based on the clinical data. The predictive attribute categories include key points of basic physiological characteristics, key points of ovarian reserve function, key points of treatment intervention, and key points related to pregnancy outcome. When the total number of key points of basic physiological characteristics, key points of ovarian reserve function, key points of treatment intervention, and key points related to pregnancy outcome exceeds the first threshold, a dimension reduction improved feedforward neural network is established. The dimension reduction improved feedforward neural network takes the data of the key points of basic physiological characteristics and key points of treatment intervention as input and outputs the predicted data of key points of ovarian reserve function and key points related to pregnancy outcome. When the sum of one-quarter of the number of key points of basic physiological characteristics, key points of ovarian reserve function, key points of treatment intervention, and key points associated with pregnancy outcome is less than a first threshold, a GNN model is established. The GNN model establishes and updates the connection relationship between hierarchical nodes based on the hierarchical relationship of the cluster nodes of the key points of basic physiological characteristics, key points of ovarian reserve function, key points of treatment intervention, and key points associated with pregnancy outcome. It continuously updates each hierarchical node based on the data of the key points of basic physiological characteristics and obtains the predicted classification of key points of ovarian reserve function and key points associated with pregnancy outcome based on the updated hierarchical nodes.

[0007] Optionally, the GNN model includes: The aggregation layer is used to establish an input node graph based on the correlation of the key points of the basic physiological features, and to generate basic physiological aggregation nodes using an aggregation function. The graph is constructed using a layer to generate a directed graph by utilizing the temporal and causal relationships of the basic physiological aggregation nodes, key points of ovarian reserve function, key points of treatment intervention, and key points related to pregnancy outcomes, wherein the key points of treatment intervention are regulatory nodes. The fully connected layer is used to utilize the graph to establish the attributes of key points related to ovarian reserve function and pregnancy outcome, and outputs the probability of pregnancy for key points related to ovarian reserve function and pregnancy outcome.

[0008] Optionally, obtaining clinical data from patients with early-onset ovarian insufficiency includes: Acquire virtual data of ovarian structure generated by a medical imaging system; Based on virtual data of ovarian structure, corresponding assisted reproductive strategies are determined, and clinical data are updated based on these strategies.

[0009] Optionally, determining the corresponding assisted reproductive strategy based on virtual ovarian structure data and updating clinical data based on the assisted reproductive strategy includes: Based on ovarian structure data, ovulation induction protocol, and embryo transfer operation reference table, adjust the dosage of ovulation induction drugs and the embryo transfer time window; The clinical data were updated based on the adjusted ovulation induction protocol and embryo transfer parameters.

[0010] Optionally, determining the number of predictive keypoints and the corresponding predictive attribute categories based on clinical data includes: Each predictive attribute category is determined according to the requirements of pregnancy prediction, and several predictive points are selected based on clinical data according to key points of ovarian reserve function, key points of treatment intervention, and key points related to pregnancy outcome. The hormone sensitivity index and follicle development index for each prediction point are determined based on several prediction points. The pregnancy potential at each prediction point is calculated based on the hormone sensitivity index and follicle development index, and the stability at each prediction point is also calculated. Key prediction points are selected from the prediction points based on the pregnancy potential and stability.

[0011] Optionally, several predictive points are selected for each category based on clinical data, including: The menstrual cycle is divided according to time series based on the patient's clinical data; Selecting key time points from the follicular phase, ovulation phase, and luteal phase of the menstrual cycle as crucial points for ovarian reserve function; The rate of change in hormone levels at each key point of ovarian reserve function was calculated using a hormone kinetic model based on data from key points of basic physiological characteristics.

[0012] Optionally, the dimensionality reduction improved feedforward neural network includes: The input layer is used to convert key points of basic physiological characteristics and key points of treatment intervention, such as age, AMH level, FSH value, treatment plan, hormone level data, hormone sensitivity index and follicle development index, into multidimensional data of prediction points. Dimensionality reduction layer is used to reduce the dimensionality of the multidimensional data of the prediction points to obtain the dimensionality-reduced data of the multidimensional data of the prediction points; Multiple hidden layers are used to extract abstract features through linear transformation; There are two output layers. One output layer is used to output the pregnancy prediction result based on the abstract features and the activation function. The other output layer is used to output the hormone level prediction result of the ovarian reserve function point based on the abstract features output by part of the hidden layer. The adjustment unit is used to adjust the bias matrix corresponding to the partially hidden layer based on the results calculated by the hormone kinetic model.

[0013] Furthermore, to achieve the above objectives, the present invention also provides a model-building system for predicting the pregnancy rate of early-onset ovarian insufficiency, the system comprising: The feature identification module is used to acquire clinical data of patients with early-onset ovarian insufficiency, and determine the number of predictive key points and the corresponding predictive attribute categories based on the clinical data. The predictive attribute categories include basic physiological characteristic key points, ovarian reserve function key points, treatment intervention key points, and pregnancy outcome related key points. The dimensionality reduction modeling module is used to establish a dimensionality reduction improved feedforward neural network when the total number of key points of basic physiological characteristics, key points of ovarian reserve function, key points of treatment intervention, and key points related to pregnancy outcome exceeds a first threshold. The dimensionality reduction improved feedforward neural network takes the data of the key points of basic physiological characteristics and key points of treatment intervention as input and outputs the predicted data of key points of ovarian reserve function and key points related to pregnancy outcome. The graph network construction module is used to build a GNN model when the sum of one-quarter of the number of basic physiological feature key points, ovarian reserve function key points, treatment intervention key points, and pregnancy outcome related key points is less than a first threshold. The GNN model establishes and updates the connection relationship between hierarchical nodes based on the hierarchical relationship of the cluster nodes of the basic physiological feature key points, ovarian reserve function key points, treatment intervention key points, and pregnancy outcome related key points. It continuously updates each hierarchical node based on the data of the basic physiological feature key points and obtains the predicted classification of ovarian reserve function key points and pregnancy outcome related key points based on the updated hierarchical nodes.

[0014] Furthermore, to achieve the above objectives, the present invention also provides a model building device for predicting the pregnancy rate of early-onset ovarian insufficiency. The device includes: a memory, a processor, and a model building program for predicting the pregnancy rate of early-onset ovarian insufficiency stored in the memory and executable on the processor. The model building program for predicting the pregnancy rate of early-onset ovarian insufficiency is configured to implement the steps of the model building method for predicting the pregnancy rate of early-onset ovarian insufficiency as described in any one of the above descriptions.

[0015] Furthermore, to achieve the above objectives, the present invention also provides a computer-readable storage medium storing a model-building program for predicting the pregnancy rate of early-onset ovarian insufficiency, wherein when the model-building program for predicting the pregnancy rate of early-onset ovarian insufficiency is executed by a processor, the steps of the model-building method for predicting the pregnancy rate of early-onset ovarian insufficiency as described above are implemented.

[0016] This invention provides a method for building a model to predict the pregnancy rate of patients with early-onset ovarian insufficiency. The method constructs a hybrid modeling strategy based on adaptive selection of data scale, employing dimensionality reduction improved feedforward neural networks and graph neural networks (GNNs) for different data dimension scenarios. This effectively integrates the nonlinear features and hierarchical causal relationships among multiple variables in high-dimensional clinical data, improving the prediction accuracy and model generalization ability for the pregnancy rate of patients with early-onset ovarian insufficiency. Simultaneously, by introducing a classification modeling mechanism that incorporates multiple key points such as basic physiological characteristics, ovarian reserve function, treatment intervention, and pregnancy outcome associations, the interpretability and clinical applicability of the model are enhanced, providing scientific and dynamic intelligent decision support for individualized fertility assessment and assisted reproductive technology program development. Attached Figure Description

[0017] Figure 1 This is a flowchart illustrating an embodiment of the method for establishing a model to predict the pregnancy rate in early-onset ovarian insufficiency according to the present invention. Figure 2 This is a structural block diagram of an embodiment of the model-building system for predicting pregnancy rates in early-onset ovarian insufficiency according to the present invention.

[0018] The realization of the objective, functional features and advantages of the present invention will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0019] It should be understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the invention.

[0020] Reference Figure 1 , Figure 1 This is a flowchart illustrating an embodiment of the method for establishing a model to predict the pregnancy rate of early-onset ovarian insufficiency according to the present invention.

[0021] In one embodiment, the method for establishing a model to predict the pregnancy rate of early-onset ovarian insufficiency includes: Step S100: Obtain clinical data of patients with early-onset ovarian insufficiency, and determine the number of predictive key points and the corresponding predictive attribute categories based on the clinical data. The predictive attribute categories include basic physiological characteristic key points, ovarian reserve function key points, treatment intervention key points, and pregnancy outcome related key points.

[0022] Clinical data for patients with early-onset ovarian insufficiency (POI) can be multi-dimensional medical information recorded during the diagnosis and treatment of POI patients, including structured or semi-structured data such as demographics, endocrinology, imaging, treatment history, and pregnancy outcomes. This data serves as the original input source for model construction, supporting the extraction of prediction key points and the selection of modeling strategies. In this embodiment, clinical data for POI patients can be collected through clinical information systems such as electronic medical records, laboratory test reports, ultrasound examination records, and follow-up databases. Prediction key points can be the smallest semantic units abstracted from clinical data that have discriminative significance for pregnancy rate prediction. They can be used as the basic variable units for model input and output, realizing the structured expression of clinical information. For example, prediction key points can include, but are not limited to, one or more of the following: continuous numerical indicators (such as AMH concentration), classification labels (such as whether hormone replacement therapy is received), and time-series events (such as age at menarche).

[0023] Predictive attribute categories can be high-level semantic groupings of predictive key points according to medical logic, which can be used to guide feature engineering and model architecture design, ensuring clinical interpretability. Furthermore, predictive attribute categories can include, but are not limited to, basic physiological characteristic key points, ovarian reserve function key points, treatment intervention key points, pregnancy outcome-related key points, etc. Basic physiological characteristic key points can be immutable or slowly changing clinical indicators reflecting the patient's inherent biological state, which can be used as initial input variables for the model, providing individual baseline information. In an exemplary embodiment, basic physiological characteristic key points can include, but are not limited to, one or more of age, age at menarche, BMI, positive family history markers, etc.

[0024] Key indicators of ovarian reserve function can be dynamic physiological indicators characterizing the quantity and quality of remaining follicles in the ovary. These can be used as intermediate predictive targets or model outputs to reflect current fertility potential. For example, key indicators of ovarian reserve function may include, but are not limited to, AMH levels, basal FSH levels, antral follicle count (AFC), and inhibin B concentration. Key indicators of treatment intervention can be clinical variables recording the medical interventions and their parameters received by the patient. These can be used as model inputs to assess the impact of different treatment regimens on pregnancy outcomes. In one specific embodiment, key indicators of treatment intervention may include, but are not limited to, whether DHEA supplementation is used, the type of hormone replacement therapy regimen, the number of ART cycles, and the type of ovulation induction drugs. Key indicators associated with pregnancy outcomes can be clinical endpoints or proxy indicators that directly or indirectly reflect pregnancy success. These can be used as the final predictive target of the model to assess the degree of fertility achievement. Furthermore, key indicators associated with pregnancy outcomes may include, but are not limited to, natural pregnancy occurrence, clinical pregnancy rate, live birth rate, and number of embryo implantation failures.

[0025] Obtaining clinical data for patients with early-onset ovarian insufficiency (POI) can be achieved by extracting structured medical records from hospital information systems that meet the POI diagnostic criteria. Furthermore, this clinical data can be collected from electronic medical record systems, laboratory test reports, ultrasound examination records, and follow-up databases, providing a raw data foundation for subsequent keypoint extraction. Determining the number of predicted keypoints and their corresponding predicted attribute categories based on clinical data can be achieved through medical expert annotation and natural language processing techniques, mapping the raw data to a predefined four-category keypoint system. Further, determining the number of predicted keypoints and their corresponding predicted attribute categories based on clinical data can be accomplished by combining structured field matching with unstructured text extraction, achieving standardization and structuring of clinical information and supporting the selection of modeling strategies.

[0026] In step S200, when the total number of key points of basic physiological characteristics, key points of ovarian reserve function, key points of treatment intervention, and key points related to pregnancy outcome exceeds the first threshold, a dimension reduction improved feedforward neural network is established. The dimension reduction improved feedforward neural network takes the data of key points of basic physiological characteristics and key points of treatment intervention as input and outputs the predicted data of key points of ovarian reserve function and key points related to pregnancy outcome.

[0027] The first threshold can be a preset critical value for the number of key points used to distinguish between high-dimensional and low-dimensional data scenarios, and can be used to trigger the model type selection mechanism in the hybrid modeling strategy. In this embodiment, the first threshold can be optimized and determined on the training set through cross-validation to balance the relationship between model complexity and sample size. The dimensionality reduction improved feedforward neural network can be a deep learning architecture that embeds a dimensionality reduction module before a standard feedforward neural network. It can be used to compress the input space in high-dimensional key point scenarios, retain nonlinear discriminative features, and alleviate overfitting. Furthermore, the dimensionality reduction improved feedforward neural network can connect dimensionality reduction components such as autoencoders, principal component analysis (PCA), or linear discriminant analysis (LDA) after the input layer, and then connect to a fully connected hidden layer. In a specific embodiment, the dimensionality reduction improved feedforward neural network can receive basic physiological feature key points and treatment intervention key points as input, and output predicted data of ovarian reserve function key points and pregnancy outcome related key points. The predicted data can be continuous or probabilistic prediction results output by the dimensionality reduction improved feedforward neural network, which can be used to quantify the expected values ​​of ovarian reserve function key points and pregnancy outcome related key points. For example, the prediction data may include, but is not limited to, AMH prediction values, pregnancy probability scores, AFC estimates, etc.

[0028] When the total number of key points exceeds a first threshold, an improved feedforward neural network with dimensionality reduction is established. This can be achieved by constructing an FNN architecture containing a dimensionality reduction module when the total number of key points in the four categories exceeds a preset threshold. Furthermore, this improved feedforward neural network can be established by adding an autoencoder after the input layer for nonlinear dimensionality reduction, followed by a three-layer fully connected network, or by using PCA to project the input to the principal component space before inputting it into a standard multilayer perceptron. This achieves the technical effect of reducing model complexity, preserving nonlinear features, and avoiding overfitting in high-dimensional scenarios.

[0029] Using data on key points of basic physiological characteristics and key points of treatment intervention as input, the numerical vectors of the two types of key points can be concatenated into a model input tensor. Furthermore, using these data as input can be standardized to form a uniform-dimensional input vector, thus achieving the technical effect of integrating patient baseline status and intervention information as a starting point for prediction. The output of predicted data on key points of ovarian reserve function and pregnancy outcome-related key points can be the regression values ​​or probabilities of the corresponding key points output from the last layer of the model. Furthermore, the output of predicted data on key points of ovarian reserve function and pregnancy outcome-related key points can be used to regress AMH, AFC, and pregnancy probability separately through a multi-task output head, thus achieving the technical effect of simultaneously predicting intermediate physiological states and final pregnancy outcomes, forming an end-to-end assessment chain.

[0030] Step S300: When the sum of one-quarter of the number of basic physiological characteristic key points, key points of ovarian reserve function, key points of treatment intervention, and key points associated with pregnancy outcome is less than a first threshold, a GNN model is established. The GNN model establishes and updates the connection relationship between hierarchical nodes based on the hierarchical relationship of cluster nodes of basic physiological characteristic key points, key points of ovarian reserve function, key points of treatment intervention, and key points associated with pregnancy outcome. Each hierarchical node is continuously updated based on the data of basic physiological characteristic key points, and the predicted classification of key points of ovarian reserve function and key points associated with pregnancy outcome is obtained based on the updated hierarchical nodes.

[0031] In this system, clustering nodes can be representative graph nodes generated from key points of basic physiological features using an unsupervised clustering algorithm. These nodes can be used as initial input nodes in the GNN, reducing the impact of sparsity in the original features. In this embodiment, clustering nodes can group key points of basic physiological features using K-means, hierarchical clustering, or spectral clustering, generating a central node for each group. Hierarchical relationships can be directed dependency structures between key point categories defined based on medical causal logic. These relationships can guide the topology construction of node connections in the GNN, ensuring that information flow conforms to the clinical reasoning path. For example, hierarchical relationships can include, but are not limited to, basic physiology → ovarian reserve, ovarian reserve → pregnancy outcome, and treatment intervention → ovarian reserve. Hierarchical nodes can be sets of graph nodes corresponding to different predicted attribute categories in the GNN. These nodes can carry category-level information and achieve cross-level feature fusion during message passing. Furthermore, hierarchical nodes can receive information from clustering nodes and interact with and update other hierarchical nodes through hierarchical relationships.

[0032] Generative Neural Networks (GNNs) are neural network models that represent hierarchical causal relationships between variables based on graph structures. They can be used to explicitly model medical logic paths in low-dimensional, sparse data scenarios, improving interpretability and small-sample generalization ability. In an exemplary embodiment, the GNN model can map key points to graph nodes, construct edge connections based on medical knowledge, and update node representations through message passing mechanisms. Furthermore, the GNN model can initialize the graph structure using clustered nodes of basic physiological feature key points, dynamically update nodes at each level through hierarchical relationships, and ultimately generate a predicted classification. The predicted classification can be a discrete category label or classification probability distribution output by the GNN model, which can be used to qualitatively judge key points associated with pregnancy outcomes, supporting clinical decision-making. For example, the predicted classification can include, but is not limited to, binary classification of pregnancy success / failure, triadic classification of low / medium / high pregnancy potential, and a natural pregnancy probability level.

[0033] A Generative Neural Network (GNN) model is established when the sum of one-quarter of the number of basic physiological feature key points and the number of other key points is less than a first threshold. This can be achieved by calculating whether (number of basic physiological feature key points / 4 + number of other three types of key points) is below the first threshold; if so, a GNN is constructed. Furthermore, establishing a GNN model when the sum of one-quarter of the number of basic physiological feature key points and the number of other key points is less than the first threshold can be achieved by constructing a heterogeneous graph, where node types correspond to the four types of key points, edges are defined according to medical guidelines, or basic physiological features are clustered into several nodes, and the remaining key points are treated as independent nodes, connected by edges according to hierarchical relationships. This achieves the technical effect of introducing prior knowledge using graph structures in data-sparse scenarios, enhancing the robustness of the model. The connection relationships between updated hierarchical nodes are established based on the hierarchical relationships of clustered nodes of basic physiological feature key points, key points of ovarian reserve function, key points of treatment intervention, and key points related to pregnancy outcomes. This can be achieved by creating directed edges in the graph according to a pre-defined medical causal path (e.g., basic → reserve → outcome). Furthermore, this operation can automatically construct edge connections through knowledge graphs or clinical guideline-driven rule engines, thereby achieving the technical effect of constructing information dissemination paths that conform to clinical logic and improving model interpretability.

[0034] Continuously updating each level of nodes based on key points of basic physiological characteristics can be achieved through the message passing mechanism of a Generative Neural Network (GNN), aggregating neighbor information layer by layer from the cluster node to update the node representation. Furthermore, this continuous updating of each level of nodes can be accomplished using a GraphSAGE aggregator to perform mean pooling updates on the neighbors of each level of node, or using a GAT attention mechanism to assign weights to neighbors from different sources and then perform weighted summation. This achieves cross-level feature fusion and dynamically reflects the interactions between variables. The predicted classification of key points related to ovarian reserve function and pregnancy outcome based on the updated level of nodes can be achieved by inputting the final embedding vector of the corresponding level node into a classification head to output the prediction result. Furthermore, this operation can be implemented using a Softmax classifier or a logistic regression head, thereby generating interpretable classification predictions and supporting clinical decision-making.

[0035] Taking individualized fertility assessment of POI patients as an example, the model establishment method for predicting the pregnancy rate of early-onset ovarian insufficiency in this embodiment can be as follows: A 32-year-old POI patient presents with clinical data including age, AMH = 0.3 ng / mL, AFC = 3, DHEA treatment for 3 months, FSH = 28 IU / L, etc. The system extracts 12 prediction key points, the total of which is greater than the first threshold (set to 10). It automatically calls the dimensionality reduction improved feedforward neural network, using age, BMI, DHEA use, etc. as basic physiological and treatment intervention inputs, and predicts that her AMH, AFC, and natural pregnancy probability are 0.28, 2.9, and 18%, respectively. Another 28-year-old patient only provides 4 key points: age, age at menarche, family history, and whether she is pregnant. The sum is less than the threshold. The system constructs a GNN: the basic physiological characteristics are clustered into 1 node, connected according to the "basic → reserve → outcome" hierarchy, and outputs "low pregnancy potential" after message passing, suggesting that ART with donor eggs should be considered as soon as possible.

[0036] In one embodiment, the GNN model includes: The aggregation layer is used to build an input node graph based on the correlation of key points of basic physiological features, and to generate basic physiological aggregation nodes using aggregation functions; The graph is constructed using layers to generate a directed graph by utilizing the temporal and causal relationships of basic physiological aggregation nodes, key points of ovarian reserve function, key points of treatment intervention, and key points related to pregnancy outcomes. Among these, the key points of treatment intervention are regulatory nodes. The fully connected layer is used to utilize the graph to establish the attributes of key points related to ovarian reserve function and pregnancy outcome, and outputs the probability of pregnancy for key points related to ovarian reserve function and pregnancy outcome.

[0037] The aggregation layer can be a submodule in the GNN model used for association modeling and information compression of key points of basic physiological characteristics. It can be used to generate representative basic physiological aggregate nodes, reduce input dimensionality, and retain key physiological representations. In this embodiment, the aggregation layer can construct an input node graph based on the intrinsic correlation of key points of basic physiological characteristics and fuse neighboring node information through an aggregation function. For example, the aggregation layer can work in conjunction with the graph building layer, and its output basic physiological aggregate nodes are used by the graph building layer as the starting nodes for constructing a directed graph, forming a hierarchical causal structure together with the other three types of key points. The correlation of key points of basic physiological characteristics can be a statistical correlation or medical logical dependency between key points of basic physiological characteristics, which can be used to guide the construction of edge connections in the input node graph, ensuring that the aggregation process conforms to clinical practice. Furthermore, the correlation of key points of basic physiological characteristics can include, but is not limited to, one or more of the following: collinearity between demographic indicators (such as age and age of menarche), correlation between metabolic indicators (such as BMI and insulin resistance markers), and consistency of genetic background (such as family history and gene polymorphism).

[0038] The input node graph can be an undirected graph structure consisting of basic physiological feature key points as nodes and their relationships as edges. It can be used as the initial graph representation for the aggregation layer, supporting subsequent node aggregation operations. In an exemplary embodiment, the input node graph can map each basic physiological feature key point to a graph node, establishing connections between nodes based on preset association rules or data-driven methods. The aggregation function can be a mathematical operation used to fuse node and neighbor information to generate a new node representation, and can be used to achieve information compression and feature extraction of the local graph structure. For example, the aggregation function can include, but is not limited to, one or more of the following: mean aggregation, max pooling aggregation, and attention-weighted aggregation.

[0039] Basic physiological aggregation nodes can be graph nodes output from the aggregation layer, representing the comprehensive semantics of a set of key basic physiological features. These nodes can be used as input to the graph construction layer, replacing the original high-dimensional basic physiological features and reducing redundancy. The input node graph is constructed based on the correlations of the key basic physiological features. This can be achieved by using the key basic physiological features as nodes and establishing undirected edge connections based on their statistical or medical correlations. Furthermore, this operation can be implemented by constructing edges based on predefined association rules based on expert knowledge (e.g., a negative correlation between age and menarche age), or by automatically establishing connections by calculating the Pearson correlation coefficient or mutual information between features and setting thresholds. This allows for the construction of an initial graph structure that conforms to physiological logic, providing a reasonable topological foundation for subsequent aggregation.

[0040] Generating basic physiological aggregate nodes using aggregation functions can be achieved by applying the aggregation function to each node and its neighbors in the input node graph to generate compressed representative nodes. In a specific embodiment, this operation can be achieved by averaging the k-hop neighbors of each basic physiological feature node using GraphSAGE-style mean aggregation, or by using the GAT mechanism to assign attention weights to different neighbors and then weighted summing to generate aggregate nodes. This reduces the dimensionality of basic physiological features, eliminates redundancy, and retains discriminative comprehensive physiological representations.

[0041] The graph construction layer can be a submodule in a GNN model used to construct a directed graph containing multiple key points and their medical logical relationships. It can be used to explicitly encode the temporal evolution and causal dependencies between variables, forming an interpretable graph topology. In this embodiment, the graph construction layer can start from basic physiological aggregation nodes, combine key points of ovarian reserve function, key points of treatment intervention, and key points related to pregnancy outcome, and construct directed edges based on medical knowledge. Temporal sequence and causal relationship can be the chronological order of clinical variables in the time dimension and the causal influence path in medicine, which can be used to guide the setting of the direction of edges in the directed graph to ensure that the information flow conforms to the actual diagnosis and treatment logic. Furthermore, temporal sequence and causal relationship can include, but are not limited to, one or more of the following: the time constraint that treatment intervention occurs after ovarian function assessment, the causal path that determines the level of ovarian reserve based on basic physiological state, and the regulatory effect of treatment measures on pregnancy outcome.

[0042] A directed graph can be a graph structure where nodes represent key points and edges represent temporal or causal relationships, and it has a direction. It can be used as a message carrier in a Generative Neural Network (GNN) to support information dissemination that conforms to clinical logic. For example, a directed graph can be generated by a graph construction layer, allowing subsequent fully connected layers to extract node attributes for prediction. A moderating direction can be a type of edge direction designated as having a regulatory effect in a directed graph, specifically referring to the influence path of a treatment intervention key point on downstream variables. This can be used to highlight the active moderating role of treatment intervention in the causal chain and enhance the model's sensitivity to the intervention effect. In an exemplary embodiment, a moderating direction can be achieved by marking edges pointing from treatment intervention key points to ovarian reserve function or pregnancy outcome-related key points as moderating directions when constructing the directed graph. Generating a directed graph using the temporal and causal relationships of basic physiological aggregation nodes, ovarian reserve function key points, treatment intervention key points, and pregnancy outcome-related key points can involve mapping these four types of elements to graph nodes and setting directed edges according to medical temporal and causal logic. Furthermore, this operation can be implemented by establishing causal edges along the main path of "basic physiology → ovarian reserve → pregnancy outcome", with treatment intervention nodes connecting both ovarian reserve and pregnancy outcome, or by introducing timestamp constraints that only allow treatment intervention nodes to point to subsequent assessment or outcome nodes. This allows the construction of a clinically interpretable directed graph structure that explicitly expresses the medical logic between variables.

[0043] Setting key points of treatment intervention as regulatory directions can be achieved by labeling edges originating from these key points in a directed graph as special types with regulatory effects. This emphasizes the active impact of treatment measures on downstream variables, assigning them higher weights or specific propagation rules during message passing. The fully connected layer can be a neural network layer at the end of a GNN model used to map graph node attributes to final predicted probabilities. It can be used to output the probability of pregnancy corresponding to key points of ovarian reserve function and key points related to pregnancy outcome. In this embodiment, the fully connected layer can receive node embedding vectors updated by the graph building layer after message passing and generate probability outputs through linear transformation and activation functions. Attributes can be numerical feature vectors obtained from each key node in the graph building layer after message passing, which can be used as input to the fully connected layer to carry fused multi-source information. For example, attributes can include, but are not limited to, one or more of node embedding vectors, graph attention weights, and aggregated physiological-therapeutic joint representations. The probability of pregnancy can be a quantitative score output by the model representing the likelihood of pregnancy for POI patients, which can be used to provide quantifiable fertility assessment results for clinical practice, supporting individualized decision-making. Furthermore, the probability of pregnancy may include, but is not limited to, one or more of the following: the probability of natural pregnancy, the probability of a successful ART cycle, and the probability of a live birth within 12 months.

[0044] By utilizing the attributes of key points related to ovarian reserve function and pregnancy outcome from the graph construction layer, the probability of pregnancy can be output. This can be achieved by inputting the updated attributes of the two types of key points from the graph construction layer after message passing into a fully connected layer, and then outputting the probability after nonlinear transformation. In a specific embodiment, this operation can be performed by concatenating the embedding vectors of ovarian reserve and pregnancy outcome nodes and inputting them into a two-layer fully connected network with sigmoid activation, or by weighting and fusing the attributes of the two types of nodes separately before inputting them into the classification head, with the weights dynamically learned through an attention mechanism. This allows for the generation of end-to-end quantitative predictions of pregnancy probability, supporting clinical decision-making.

[0045] Taking pregnancy prediction after DHEA intervention in POI patients as an example, the model building method for predicting the pregnancy rate of early-onset ovarian insufficiency in this embodiment can be as follows: A 30-year-old POI patient has basic physiological characteristics including age, age of menarche at 14 years old, BMI of 22, and no family history. The system identifies four key points of basic physiological characteristics, calculates their correlation through an aggregation layer (e.g., age is negatively correlated with age of menarche), constructs an input node graph, and generates one basic physiological aggregation node through mean aggregation. The graph building layer models this node together with AMH=0.5, AFC=4 (ovarian reserve), DHEA use for 3 months (treatment intervention), and no previous pregnancy (pregnancy outcome). A directed graph is constructed based on causal and temporal relationships such as "basic → reserve", "treatment → reserve", and "result", and the edges from the DHEA node to AMH and pregnancy outcome are marked as regulating directions. After GNN message passing, the fully connected layer outputs a probability of 22% for natural pregnancy within the next 6 months based on the updated AMH and pregnancy outcome node attributes.

[0046] In one embodiment, clinical data of patients with early-onset ovarian insufficiency are obtained, including: Obtain virtual data of ovarian structure generated by a medical imaging system.

[0047] The medical imaging system can be a medical imaging device and its post-processing software platform used to generate visualized ovarian structural data. It can provide high-resolution ovarian anatomical information to support the construction of virtual data. In this embodiment, the medical imaging system can acquire raw images using imaging technologies such as transvaginal ultrasound and pelvic MRI, and then generate structured virtual data using a three-dimensional reconstruction algorithm. Exemplarily, the medical imaging system can include, but is not limited to, one or more of a three-dimensional ultrasound imaging system, a high-field MRI system, and an optical coherence tomography (OCT) system. The virtual ovarian structural data can be a digital three-dimensional model or quantitative indicator reconstructed by the medical imaging system, characterizing the anatomical morphology and internal structure of the ovary. It can be used as a supplementary modality to clinical data, enhancing the structural assessment of ovarian reserve function. In an exemplary embodiment, the virtual ovarian structural data can include, but is not limited to, one or more of an ovarian volume reconstruction model, a antral follicle spatial distribution heatmap, and a cortex-medullum ratio quantitative indicator.

[0048] Obtaining virtual ovarian structural data generated by a medical imaging system can be achieved by extracting post-processed 3D ovarian reconstruction results or structural quantification reports from hospital PACS or ultrasound workstations. Furthermore, this operation can be automatically synchronized with structured data output from an image post-processing platform via a standardized interface, thereby incorporating anatomical information into a multimodal clinical data system and expanding the assessment dimensions of traditional hormone / count indicators.

[0049] Based on virtual data of ovarian structure, corresponding assisted reproductive strategies are determined, and clinical data are updated based on these strategies.

[0050] Assisted reproductive strategies (ART) can be ART intervention protocols such as ovulation induction, oocyte retrieval, and embryo transfer, tailored to individual patient characteristics. These protocols can serve as a refined source of key treatment intervention points, enriching the decision-making semantics of clinical data. In one specific embodiment, ART strategies may include, but are not limited to, one or more of the following: microstimulation protocols, natural cycle oocyte retrieval strategies, and donor IVF pathways. Determining the corresponding ART strategy based on virtual ovarian structure data can be achieved by matching pre-defined clinical decision rules or expert system recommendations based on structural features such as ovarian volume and follicle distribution. Furthermore, this operation can be implemented by automatically recommending microstimulation protocols based on ovarian volume <3mL and AFC <5, or by suggesting a unilateral puncture-based oocyte retrieval strategy if antral follicles are focally distributed. This allows for a mapping from structural features to treatment decisions, enhancing the individualization and interpretability of key treatment intervention points. Updating clinical data based on ART strategies can involve encoding the derived ART strategies as new key treatment intervention points and linking them to the original clinical records. Furthermore, this operation can be achieved by writing structured data into electronic medical record systems or research databases, thereby forming a closed-loop feedback of "structure-intervention-outcome" and strengthening the explicitness of the causal logic between key points.

[0051] For example, in the scenario of initial ART evaluation for POI patients, the model establishment method for predicting pregnancy rate in this embodiment could be as follows: A 30-year-old POI patient has AMH = 0.4 ng / mL, basal FSH = 25 IU / L, and routine indicators suggest low ovarian reserve. The system retrieves virtual data of the ovarian structure reconstructed by transvaginal 3D ultrasound, showing that the right ovary has a volume of only 1.8 mL but contains 3 focally distributed antral follicles. Based on this, the "natural cycle + unilateral priority oocyte retrieval" assisted reproductive strategy is determined, and this strategy is incorporated into the clinical data as a new key point of treatment intervention. Subsequent models, when predicting pregnancy rates, not only use the original hormonal indicators but also integrate this structure-driven intervention strategy to output more accurate individualized prediction results.

[0052] In one embodiment, a corresponding assisted reproductive strategy is determined based on virtual ovarian structure data, and clinical data is updated based on the assisted reproductive strategy, including: Based on ovarian structure data, ovulation induction protocol, and embryo transfer operation reference table, adjust the dosage of ovulation induction drugs and the embryo transfer time window; The clinical data were updated based on the adjusted ovulation induction protocol and embryo transfer parameters.

[0053] The ovulation induction protocol can be a gonadotropin-based drug administration plan tailored to the patient's ovarian responsiveness, serving as a core component of assisted reproductive strategies to guide follicle development regulation. In this embodiment, the ovulation induction protocol may include, but is not limited to, one or more of antagonist protocols, microstimulation protocols, and natural cycle monitoring protocols. The embryo transfer operation control table can be a standardized ART operation parameter mapping rule base based on ovarian structural characteristics and endometrial status, providing decision-making basis from anatomical / physiological data to specific transfer time windows. For example, the embryo transfer operation control table can be a structured rule table constructed from clinical guidelines, historical successful cases, and expert consensus. Further, the embryo transfer operation control table may include, but is not limited to, endometrial-follicle synchrony control tables, luteal support initiation timing control tables, and frozen embryo thawing and transfer window tables. The ovulation induction drug dosage can be the specific amount of gonadotropins (such as FSH and LH) administered per unit cycle, directly affecting the quantity and quality of follicle recruitment, and is a quantitative expression of key points in treatment intervention. In an exemplary embodiment, the ovulation induction drug dosage may include, but is not limited to, the initial dose, incremental adjustment dose, and maintenance dose.

[0054] The embryo transfer time window, determined based on follicular development and endometrial receptivity, is the optimal time for embryo implantation. It can be used to determine implantation success rates and is a crucial predictor of pregnancy outcomes. In one specific embodiment, the embryo transfer time window may include, but is not limited to, the fresh embryo transfer date (day 3 or 5 after oocyte retrieval), the frozen embryo transfer window in a natural cycle, or a fixed date after the LH surge in a natural cycle. Embryo transfer parameters can be a set of clinical variables describing the specific settings of the embryo transfer procedure. They can be used to update clinical data and refine key points of treatment intervention. Furthermore, embryo transfer parameters may include, but are not limited to, the number of embryos transferred, the embryonic developmental stage (cleavage stage / blastocyst), and the type of transfer catheter.

[0055] Based on ovarian structural data and ovulation induction protocols and embryo transfer operation tables, the dosage of ovulation induction drugs and the embryo transfer time window can be adjusted. This can be achieved by matching indicators such as volume and follicle distribution in the virtual ovarian structural data with the rules in the table, outputting individualized drug dosage and transfer time recommendations. Furthermore, this operation can be implemented by using a standard antagonist protocol (225 IU) with an initial FSH dose of 150 IU / day and setting the fresh embryo transfer time to day 5 after oocyte retrieval if the ovarian volume is <2 mL and AFC ≤3; or by using a standard antagonist protocol (225 IU) with an open day 3 / 5 dual-window assessment if the antral follicles are evenly distributed bilaterally and the endometrial thickness is ≥7 mm. This allows for precise mapping from static anatomical information to dynamic treatment parameters, improving the individualization and physiological adaptability of intervention measures. Updating clinical data based on the adjusted ovulation induction protocol and embryo transfer parameters can involve encoding the newly determined drug dosage, transfer time window, and embryo parameters as structured fields and adding them to the original clinical records. This enriches the semantic details of key points in treatment intervention and key points related to pregnancy outcomes, and strengthens the explicitness of causal logic in the data loop.

[0056] Taking the optimization of a second ART cycle for a POI patient as an example, the model establishment method for predicting the pregnancy rate in early-onset ovarian insufficiency in this embodiment can be as follows: A 34-year-old POI patient used the standard 225 IU FSH protocol in her first cycle and only retrieved 2 oocytes. Virtual data of ovarian structure showed atrophy of the left ovary (volume 1.2 mL) and 4 focal follicles on the right side. The system called the embryo transfer operation checklist, matched the "unilateral dominant follicle + low reserve" entry, automatically adjusted the ovulation induction drug dose to 150 IU and limited the puncture to the right side; at the same time, based on the endometrial growth rate, the embryo transfer time window was set to the afternoon of the 5th day after oocyte retrieval. This adjusted protocol and parameters were written into the clinical data and used as input for the subsequent pregnancy rate prediction model, enabling the model to more accurately assess the success probability of this cycle.

[0057] In one embodiment, determining the number of predicted keypoints and the corresponding predicted attribute categories based on clinical data includes: Each predictive attribute category is determined according to the requirements of pregnancy prediction, and several predictive points are selected based on clinical data according to key points of ovarian reserve function, key points of treatment intervention, and key points related to pregnancy outcome. The hormone sensitivity index and follicle development index for each prediction point are determined based on several prediction points. The pregnancy potential at each prediction point was calculated based on the hormone sensitivity index and follicle development index, and the stability of each prediction point was also calculated. Key prediction points are selected from the prediction points based on pregnancy potential and stability.

[0058] The pregnancy prediction requirements can be specific specifications for the granularity, target type, and time range of pregnancy outcome prediction in clinical or research scenarios, which can guide the selection range of prediction attribute categories and the setting of modeling goals. In this embodiment, the pregnancy prediction requirements may include, but are not limited to, whether to distinguish between natural pregnancy and ART pregnancy, whether to predict live birth rate rather than just clinical pregnancy, and whether to limit the prediction time window (e.g., within 6 months), among other things. Several prediction points can be candidate variable instances that have been initially screened from clinical data and belong to the three categories of key points, which can be used as the basic units for calculating the hormone sensitivity index and follicle development index. For example, several prediction points may include AMH value, E2 peak in the ovulation induction cycle, number of follicles on the hCG day, and type of luteal support protocol, etc.

[0059] The category of each predictive attribute is determined according to the pregnancy prediction requirements. This can be based on clinical needs (such as whether ART is included, prediction timeframe, etc.) to decide whether to enable all four categories or only some predictive attribute categories. Furthermore, this operation can be achieved by matching a preset rule base, thereby ensuring that the modeling goals are aligned with the actual application scenario and avoiding the introduction of irrelevant features. Several predictive points are selected based on clinical data according to key points of ovarian reserve function, key points of treatment intervention, and key points related to pregnancy outcome. This can be achieved by extracting quantifiable variables belonging to these three categories from the original clinical data as candidate predictive points. In an exemplary embodiment, this operation can be achieved by automatically labeling the variable categories using a structured electronic medical record system, thus focusing on dynamic variables directly related to pregnancy and excluding interference from purely static basic physiological characteristics (such as age at menarche) at this stage.

[0060] Hormone sensitivity index (HSI) is a composite indicator that quantifies a patient's response to endogenous or exogenous hormone stimulation. It can be used to reflect an individual's endocrine regulatory capacity and to assess the effectiveness of treatment interventions. In one specific embodiment, HSI can be calculated based on hormone levels such as FSH, LH, E2, and P, and their dynamic change rates, using a weighted function or machine learning regression model. Follicle development index (FDI) is a comprehensive score characterizing follicle recruitment, growth, and maturation efficiency. It can be used to measure the actual output capacity of ovarian function and connect ovulation reserve status with pregnancy outcomes. Furthermore, FDI can be constructed by integrating ultrasound and biochemical indicators such as AFC, number of dominant follicles in the cycle, follicle growth rate, and ovulation monitoring results.

[0061] Determining the hormone sensitivity index and follicle development index for each prediction point based on several prediction points can be achieved by standardizing the original indicators associated with each prediction point and then substituting them into a pre-trained index calculation model. For example, this operation can employ a multivariate linear regression model, using hormone changes and follicle responses in historical cohorts as labels to train index weights; or it can use a random forest regressor to automatically learn the contribution of nonlinear interaction terms to the index, thereby transforming discrete clinical indicators into physiologically meaningful continuous indices and enhancing cross-patient comparability.

[0062] Pregnancy potential can be a quantitative value of fertility calculated based on the fusion of hormone sensitivity index and follicle development index, and can be used as one of the core criteria for prioritizing prediction points. In this embodiment, pregnancy potential maps the hormone sensitivity index and follicle development index to a score between 0 and 1 using a preset function (such as product, weighted sum, or nonlinear combination). Stability can be an indicator that measures the degree of numerical fluctuation of the prediction point over time or during the treatment cycle, and can be used to assess the reliability and representativeness of the prediction point as a modeling input. Furthermore, stability can be calculated based on the standard deviation, coefficient of variation, or time series autocorrelation of multiple measurements.

[0063] The pregnancy potential of each predicted point is calculated based on the hormone sensitivity index and the follicle development index. This can be achieved by mapping the two indices to a single pregnancy potential score using a pre-defined fusion function. In one specific embodiment, this operation can employ a geometric mean, i.e., taking the square root of the product of the hormone sensitivity index and the follicle development index; or it can use a logistic function combination, i.e., weighting the two indices and mapping them using a sigmoid function. This allows for the synergistic evaluation of multi-dimensional physiological information, avoiding overestimation of fertility from a single dimension. The stability of each predicted point is calculated based on the observations of the same predicted point at multiple time points or cycles, calculating its statistical volatility index. For example, this operation can calculate the coefficient of variation (standard deviation divided by the mean) as a negative stability indicator; or, after fitting a linear trend, take the residual variance as a measure of instability, thus identifying predicted points that are significantly affected by random fluctuations or measurement errors, improving the robustness of key points. Key predicted points are selected from the predicted points based on pregnancy potential and stability. This can be achieved by setting a dual threshold or Pareto front screening strategy to retain predicted points with both high potential and high stability. In one exemplary embodiment, this operation can retain only prediction points with a pregnancy potential greater than 0.5 and a coefficient of variation less than 0.3; or select non-dominated solution sets as key points on the pregnancy potential-stability two-dimensional plane, thereby generating a high-quality, low-redundancy, and highly representative set of modeling inputs to optimize subsequent model performance.

[0064] Taking the pre-ART cycle assessment of POI patients as an example, the model establishment method for predicting the pregnancy rate of early-onset ovarian insufficiency in this embodiment can be as follows: A 35-year-old POI patient plans to undergo IVF. Clinical data includes three AMH tests (0.28, 0.31, 0.29), E2 peak values ​​in two stimulation cycles (1800, 2100 pg / mL), and the number of follicles on the hCG day (4, 5). Based on the pregnancy prediction requirement of "predicting the clinical pregnancy rate of ART," the system determines the activation of three attributes: ovarian reserve, treatment intervention, and pregnancy outcome. AMH, E2 peak value, and follicle count are selected as several prediction points. The calculated hormone sensitivity index (based on FSH / E2 feedback strength) for AMH is 0.42, the follicle development index is 0.38, the pregnancy potential is 0.40, and the stability (coefficient of variation = 0.05) is high; the hormone sensitivity index for E2 peak value is 0.65, the follicle development index is 0.70, the pregnancy potential is 0.67, and the stability is moderate (coefficient of variation = 0.15). Ultimately, the peak values ​​of AMH and E2 were selected as the key prediction points for subsequent dimensionality reduction FNN models because they possess both high potential and acceptable stability.

[0065] In one embodiment, several prediction points are selected according to each category based on clinical data, including: The menstrual cycle is divided according to time series based on the patient's clinical data; Selecting key time points from the follicular phase, ovulation phase, and luteal phase of the menstrual cycle as crucial points for ovarian reserve function; The rate of change in hormone levels at each key point of ovarian reserve function was calculated using a hormone kinetic model based on data from key points of basic physiological characteristics.

[0066] The time series can be a collection of clinical observation data arranged chronologically to characterize the dynamic evolution of a patient's physiological state. In this embodiment, the time series can support the segmentation of the menstrual cycle and the identification of key time nodes. For example, the time series can include, but is not limited to, one or more of hormone testing time series, ultrasound monitoring time series, and symptom recording time series. The menstrual cycle can be a complete physiological cycle from the first day of one menstrual period to the first day of the next menstrual period. Furthermore, the menstrual cycle can serve as the basic time unit for dynamic assessment of ovarian function. In an exemplary embodiment, the menstrual cycle can be a natural menstrual cycle, a drug-induced cycle, or an anovulatory cycle. Dividing the menstrual cycle according to the time series based on the patient's clinical data can be achieved by segmenting continuous clinical records into independent cycle units based on the menstrual start date, ovulation monitoring, or hormone peaks. Furthermore, this operation can be implemented by manually annotating the menstrual start date combined with an automatic event detection algorithm, thereby establishing a standardized physiological time framework and supporting the extraction of stage-specific features.

[0067] The follicular phase can be the period from the onset of menstruation to ovulation in the menstrual cycle, characterized primarily by follicle recruitment and development. In one specific embodiment, the follicular phase can provide a time window for assessing the responsiveness of the basal follicle pool. Exemplarily, the follicular phase may include, but is not limited to, the early follicular phase (days 2-5 of menstruation), the mid-follicular phase (days 6-10 of menstruation), and the late follicular phase (within 3 days before ovulation). Ovulation can be the brief window during the menstrual cycle when the follicle ruptures and releases the oocyte. Further, ovulation can reflect a key point in the coordination of the hypothalamus-pituitary-ovarian axis. In one exemplary embodiment, ovulation may include, but is not limited to, the rising limb of the LH surge, the peak value of the LH surge, and the moment of ovulation. The luteal phase can be the period from ovulation to the next menstrual period, characterized by corpus luteum formation and progesterone secretion. In this embodiment, the luteal phase can be a time period for assessing corpus luteum function and endometrial receptivity. For example, the luteal phase may include, but is not limited to, the early luteal phase (1-3 days after ovulation), the mid-luteal phase (4-7 days after ovulation), and the late luteal phase (8-14 days after ovulation). The time points can be specific representative time points or events selected from various stages of the menstrual cycle. Furthermore, the time points can serve as carriers of key points in ovarian reserve function, carrying temporal hormonal information. In a specific embodiment, the time points may include, but are not limited to, the day of the lowest FSH level, the day of the rapid rise in E2, the day of the LH peak, and the first day of progesterone > 3 ng / mL.

[0068] Selecting key points for ovarian reserve function from the follicular phase, ovulation phase, and luteal phase of the menstrual cycle can be achieved by choosing physiologically significant or data-rich moments within each phase. Furthermore, this approach can be implemented by automatically identifying the starting point of E2 acceleration as a key point in the follicular phase using a hormone inflection point detection algorithm, or by manually marking the LH peak day as a key point in the ovulation phase and combining this with ultrasound confirmation of ovulation. This transforms static ovarian reserve indicators into phase-specific dynamic observation points, enhancing the expression of time-dimensional information.

[0069] Hormone kinetic models can be mathematical or computational models describing the changes in hormone concentration over time, constructed based on endocrine feedback mechanisms. In this embodiment, the hormone kinetic model can deduce the dynamic rate of hormone change at each time point from basic physiological characteristics. Furthermore, the hormone kinetic model can employ a system of ordinary differential equations (such as a Goodwin oscillator) or a data-driven recurrent neural network, combined with individual baseline parameters (age, BMI, basal FSH, etc.) for personalized fitting. The rate of change in hormone levels can be the slope or derivative of the change in a specific hormone concentration per unit time, reflecting the response speed of the endocrine system. In an exemplary embodiment, the rate of change in hormone levels can serve as a dynamic feature of key points in ovarian reserve function, capturing abnormal hormone fluctuation patterns in POI patients. Exemplary examples include, but are not limited to, the rate of increase in FSH, the slope of E2 growth, the steepness of the LH peak, and the rate of decline in progesterone.

[0070] Hormone level change rates at each key point of ovarian reserve function are calculated using a hormone kinetic model based on data from key points of basic physiological characteristics. This can be achieved by inputting basic physiological characteristics such as age, BMI, and baseline FSH into a personalized hormone kinetic model, and outputting hormone derivative values ​​at each key point. Furthermore, this operation can be achieved by using a parameterized ODE model and calibrating individual parameters through Bayesian inference to solve for local derivatives, or by using an LSTM-based surrogate model to generate hormone time series based on basic characteristics and performing numerical differentiation. This allows for the generation of dynamic features with physiological mechanism interpretation, replacing the original static hormone values ​​and improving the model's sensitivity to POI pathological patterns.

[0071] For example, in the scenario of monitoring the natural cycle of POI patients, the model establishment method for predicting the pregnancy rate of early-onset ovarian insufficiency in this embodiment can be as follows: A 31-year-old POI patient provides 6 menstrual cycle records, including daily basal body temperature, AMH, FSH, and E2 tests twice a month, and 3 vaginal ultrasounds. The system first divides the cycle into 6 periods according to the menstrual start date; in each cycle, the day when E2 begins to rise rapidly during the follicular phase (e.g., day 7 of the cycle), the day when LH peaks during ovulation (day 14 of the cycle), and the day when progesterone peaks during the luteal phase (day 21 of the cycle) are selected as time nodes; then, the patient's basic physiological characteristics, such as age 31, BMI 22.5, and average basal FSH 25 IU / L, are input into the pre-trained hormone dynamics ODE model, and the E2 change rate on the day of E2 rise is calculated to be +18 pg / mL / day (lower than normal >30), and the LH rise rate on the day of LH peak is +15 IU / L / 12h (significantly slow). These rates of change were encoded as dynamic features corresponding to key points of ovarian reserve function, and input into subsequent GNN models to reveal the pathological features of delayed follicle development initiation and weak ovulation triggering, ultimately predicting a natural pregnancy probability of <5%.

[0072] In one embodiment, the dimensionality reduction improvement of the feedforward neural network includes: The input layer is used to convert key points of basic physiological characteristics and key points of treatment intervention, such as age, AMH level, FSH value, treatment plan, hormone level data, hormone sensitivity index and follicle development index, into multidimensional data of prediction points.

[0073] The input layer can be a structural layer in a neural network that receives and initially encodes the original feature vectors. It can be used to uniformly convert multi-source clinical variables into numerical tensors that the model can process. In an exemplary embodiment, the input layer can standardize continuous variables such as age and AMH levels, and embed or one-hot encode categorical variables such as treatment regimens. Furthermore, the input layer can form a data generation relationship with the multidimensional data of prediction points. For example, the multidimensional data of prediction points output by the input layer can include, but is not limited to, one or more of the following: original hormone index vectors, derived index vectors, and treatment encoding vectors.

[0074] Converting key physiological characteristics and treatment interventions—including age, AMH levels, FSH values, treatment regimens, hormone levels, hormone sensitivity indices, and follicle development indices—into multidimensional predictive data can be achieved by standardizing, encoding, and concatenating various variables to form a unified input vector. Furthermore, this operation can be implemented by performing Z-score standardization on continuous variables and using embedding layers or one-hot encoding on categorical variables, thereby enabling the structured fusion of multi-source heterogeneous clinical data and providing high-quality input for subsequent dimensionality reduction.

[0075] The dimensionality reduction layer is used to reduce the dimensionality of the multidimensional data of the prediction points, resulting in dimensionality-reduced data of the multidimensional data of the prediction points.

[0076] The dimensionality reduction layer can be a feature compression module located after the input layer and before the hidden layer. It can be used to reduce the input dimensionality, alleviate redundancy and overfitting, and retain nonlinear discriminative information. In a specific embodiment, the dimensionality reduction layer can use linear projection (such as a PCA weight matrix) or a nonlinear autoencoder structure to achieve dimensionality compression. Furthermore, the dimensionality reduction layer can form an input-output mapping relationship with the dimensionality-reduced data. For example, the dimensionality-reduced data can include, but is not limited to, one or more of the following: principal component score vectors, autoencoder bottleneck layer outputs, and discriminative latent variables. Dimensionality reduction of the multidimensional data of prediction points can be achieved by compressing the feature dimension using a learnable projection matrix in the dimensionality reduction layer or a preset dimensionality reduction algorithm. Further, this operation can be achieved by using a single-layer autoencoder with ReLU activation as the dimensionality reduction layer, or by using a fixed PCA principal component loading matrix for linear projection, thereby reducing model complexity, suppressing noise and redundancy, and improving small-sample generalization ability.

[0077] Multiple hidden layers are used to extract abstract features through linear transformations.

[0078] The multiple hidden layers can be a sequence of fully connected neural network layers that perform nonlinear feature abstraction, which can be used to extract high-order combined features layer by layer to model complex interactions between variables. In an exemplary embodiment, the multiple hidden layers can sequentially include a shallow feature combination layer, a mid-level semantic abstraction layer, and a deep decision representation layer. Furthermore, the multiple hidden layers can form a feature generation path with the abstract features. For example, the abstract features can include, but are not limited to, one or more of hormone-treatment interaction features, reserve-outcome coupling features, and individual response pattern encoding. Extracting abstract features through linear transformation can be achieved by sequentially performing weight matrix multiplication and nonlinear activation in the multiple hidden layers. Furthermore, this operation can be implemented by appending a nonlinear activation function such as ReLU or Swish to each layer, thereby constructing high-order feature representations layer by layer and capturing nonlinear interactions between variables.

[0079] There are two output layers. One output layer is used to output the pregnancy prediction result based on the abstract features and the activation function. The other output layer is used to output the hormone level prediction result of the ovarian reserve function point based on the abstract features output by part of the hidden layer.

[0080] The two output layers can be independent output heads for pregnancy outcome prediction and ovarian reserve hormone level prediction, respectively, which can be used to achieve multi-task learning and enhance the physiological consistency within the model. In a specific embodiment, the two output layers can include a pregnancy probability output head and a hormone level regression output head. Furthermore, the pregnancy prediction result can be a quantitative value (probability or classification label) of the likelihood of pregnancy output by the model, which can be used to directly serve clinical fertility assessment decisions. For example, the pregnancy prediction result can include, but is not limited to, one or more of the following: natural pregnancy probability, ART clinical pregnancy probability, and live birth probability score. Simultaneously, the hormone level prediction result for ovarian reserve function points can be an estimate of ovarian reserve-related hormones such as AMH, FSH, and E2 at key time points, which can be used to provide intermediate physiological state supervision signals and improve the biological rationality of the model. For example, the hormone level prediction result for ovarian reserve function points can include, but is not limited to, one or more of the following: predicted AMH concentration, predicted FSH on day 3 of the cycle, and predicted pre-ovulatory E2 peak. The pregnancy prediction result is output using an activation function based on abstract features, which can be achieved by inputting deep abstract features into the output layer activated by sigmoid or softmax. Furthermore, this operation can be achieved by configuring a sigmoid function at the end of the output layer to output a probability value between 0 and 1, thereby generating interpretable probabilistic pregnancy predictions to support clinical decision-making.

[0081] Based on the abstract features output from some hidden layers, the model can predict hormone levels for ovarian reserve function points. This can be achieved by branching out a regression head from an intermediate hidden layer to predict specific hormone values. Furthermore, this operation can be implemented by deriving linear regression from the penultimate hidden layer to output AMH and FSH, or by using independent subnetworks to dynamically predict E2 based on shallow features. This allows the model to learn representations consistent with endocrine patterns through task-constrained learning, enhancing physiological consistency.

[0082] The adjustment unit is used to adjust the bias matrix corresponding to some hidden layers based on the results calculated by the hormone kinetic model.

[0083] The adjustment unit can be a regulatory module that dynamically corrects the internal parameters of the network based on the output of an external mechanism model, and can be used to inject prior knowledge of hormone kinetics into the neural network training process. In an exemplary embodiment, the adjustment unit can apply regularization or direct correction to a specified hidden layer bias term based on the theoretical rate of change calculated by the hormone kinetic model, outside of backpropagation. Furthermore, the bias matrix can be a set of offset parameters for each neuron in the hidden layer of the neural network, which can be used to influence the neuron activation threshold; the adjustment unit introduces physiological constraints by modifying its values. Exemplarily, the bias matrix can be one or more of the following, including but not limited to shallow bias vectors, mid-layer bias matrices, and task-specific bias sets.

[0084] Adjusting the bias matrices of some hidden layers based on the results calculated by the hormone kinetic model can be achieved by using the theoretical hormone change rate output by the hormone kinetic model as a priori to correct the corresponding hidden layer bias terms. Furthermore, this operation can be implemented by adding an L2 regularization term to the loss function to account for the bias term and the deviation from the kinetic prediction, or by fine-tuning the bias matrix according to the kinetic gradient direction after each forward propagation. This allows mechanism-driven knowledge to be embedded into the data-driven model, preventing the network from deviating from known physiological patterns in small samples and improving interpretability and robustness.

[0085] Taking the prediction of IVF cycle pre-pregnancy rate in POI patients as an example, the model establishment method for predicting the pregnancy rate of POI patients in this embodiment can be as follows: Input data for a 34-year-old POI patient includes: age 34, AMH = 0.25 ng / mL, basal FSH = 26 IU / L, DHEA + estrogen pretreatment, hormone sensitivity index 0.48, and follicle development index 0.41. The input layer encodes this data into a 12-dimensional vector; the dimensionality reduction layer compresses the data to 5 dimensions using an autoencoder; after extracting abstract features through three hidden layers, one output layer predicts the clinical pregnancy probability of IVF to be 22%, and another output layer simultaneously predicts the daily E2 level of hCG to be 1900 pg / mL. Meanwhile, the external hormone dynamics (ODE) model simulates an E2 rise slope of +15 pg / mL / day based on the patient's parameters, while the current network's implicit E2 dynamic trend shows +25 pg / mL / day, indicating a bias. The adjustment unit applies a negative correction to the bias matrix of the second hidden layer, making subsequent training more closely resemble physiological reality. The final model output conforms to both the data distribution and the endocrine mechanism, improving clinical reliability.

[0086] In addition, refer to Figure 2 To achieve the above objectives, the present invention also provides a model-building system for predicting the pregnancy rate of early-onset ovarian insufficiency, the system comprising: The feature labeling module 10 is used to acquire clinical data of patients with early-onset ovarian insufficiency, and determine the number of prediction key points and the corresponding prediction attribute categories based on the clinical data. The prediction attribute categories include basic physiological characteristic key points, ovarian reserve function key points, treatment intervention key points, and pregnancy outcome related key points. The dimensionality reduction modeling module 20 is used to establish a dimensionality reduction improved feedforward neural network when the total number of basic physiological feature key points, ovarian reserve function key points, treatment intervention key points and pregnancy outcome related key points is greater than a first threshold. The dimensionality reduction improved feedforward neural network takes the data of the basic physiological feature key points and treatment intervention key points as input and outputs the predicted data of ovarian reserve function key points and pregnancy outcome related key points. The graph network construction module 30 is used to establish a GNN model when the sum of one-quarter of the number of basic physiological feature key points, ovarian reserve function key points, treatment intervention key points, and pregnancy outcome related key points is less than a first threshold. The GNN model establishes and updates the connection relationship between hierarchical nodes based on the hierarchical relationship of the cluster nodes of the basic physiological feature key points, ovarian reserve function key points, treatment intervention key points, and pregnancy outcome related key points. It continuously updates each hierarchical node based on the data of the basic physiological feature key points and obtains the predicted classification of ovarian reserve function key points and pregnancy outcome related key points based on the updated hierarchical nodes.

[0087] Other embodiments or specific implementations of the model establishment system for predicting pregnancy rates in early-onset ovarian insufficiency described in this invention can be found in the above-described method embodiments, and will not be repeated here.

[0088] Furthermore, to achieve the above objectives, the present invention also provides a model building device for predicting the pregnancy rate of early-onset ovarian insufficiency. The device includes: a memory, a processor, and a model building program for predicting the pregnancy rate of early-onset ovarian insufficiency stored in the memory and executable on the processor. The model building program for predicting the pregnancy rate of early-onset ovarian insufficiency is configured to implement the steps of the model building method for predicting the pregnancy rate of early-onset ovarian insufficiency as described in any one of the above descriptions.

[0089] Furthermore, to achieve the above objectives, the present invention also provides a computer-readable storage medium storing a model-building program for predicting the pregnancy rate of early-onset ovarian insufficiency, wherein when the model-building program for predicting the pregnancy rate of early-onset ovarian insufficiency is executed by a processor, the steps of the model-building method for predicting the pregnancy rate of early-onset ovarian insufficiency as described above are implemented.

[0090] The above are merely preferred embodiments of the present invention and do not limit the scope of the patent. Any equivalent structural or procedural transformations made based on the description and drawings of the present invention, or direct or indirect applications in other related technical fields, are similarly included within the scope of patent protection of the present invention.

Claims

1. A method for establishing a model to predict the pregnancy rate in patients with early-onset ovarian insufficiency, characterized in that, The method includes: Acquire clinical data of patients with early-onset ovarian insufficiency, and determine the number of predictive key points and the corresponding predictive attribute categories based on the clinical data. The predictive attribute categories include key points of basic physiological characteristics, key points of ovarian reserve function, key points of treatment intervention, and key points related to pregnancy outcome. When the total number of key points of basic physiological characteristics, key points of ovarian reserve function, key points of treatment intervention, and key points related to pregnancy outcome exceeds the first threshold, a dimension reduction improved feedforward neural network is established. The dimension reduction improved feedforward neural network takes the data of the key points of basic physiological characteristics and key points of treatment intervention as input and outputs the predicted data of key points of ovarian reserve function and key points related to pregnancy outcome. When the sum of one-quarter of the number of key points of basic physiological characteristics, key points of ovarian reserve function, key points of treatment intervention, and key points associated with pregnancy outcome is less than a first threshold, a GNN model is established. The GNN model establishes and updates the connection relationship between hierarchical nodes based on the hierarchical relationship of the cluster nodes of the key points of basic physiological characteristics, key points of ovarian reserve function, key points of treatment intervention, and key points associated with pregnancy outcome. It continuously updates each hierarchical node based on the data of the key points of basic physiological characteristics and obtains the predicted classification of key points of ovarian reserve function and key points associated with pregnancy outcome based on the updated hierarchical nodes.

2. The method for establishing a model to predict the pregnancy rate of early-onset ovarian insufficiency as described in claim 1, characterized in that, The GNN model includes: The aggregation layer is used to establish an input node graph based on the correlation of the key points of the basic physiological features, and to generate basic physiological aggregation nodes using an aggregation function. The graph is constructed using a layer to generate a directed graph by utilizing the temporal and causal relationships of the basic physiological aggregation nodes, key points of ovarian reserve function, key points of treatment intervention, and key points related to pregnancy outcomes, wherein the key points of treatment intervention are regulatory nodes. The fully connected layer is used to utilize the graph to establish the attributes of key points related to ovarian reserve function and pregnancy outcome, and outputs the probability of pregnancy for key points related to ovarian reserve function and pregnancy outcome.

3. The method for establishing a model to predict the pregnancy rate of early-onset ovarian insufficiency as described in claim 1, characterized in that, The acquisition of clinical data from patients with early-onset ovarian insufficiency includes: Acquire virtual data of ovarian structure generated by a medical imaging system; Based on virtual data of ovarian structure, corresponding assisted reproductive strategies are determined, and clinical data are updated based on these strategies.

4. The method for establishing a model to predict the pregnancy rate of early-onset ovarian insufficiency as described in claim 3, characterized in that, The process of determining the corresponding assisted reproductive strategy based on virtual ovarian structure data and updating clinical data based on the assisted reproductive strategy includes: Based on ovarian structure data, ovulation induction protocol, and embryo transfer operation reference table, adjust the dosage of ovulation induction drugs and the embryo transfer time window; The clinical data were updated based on the adjusted ovulation induction protocol and embryo transfer parameters.

5. The method for establishing a model to predict the pregnancy rate of early-onset ovarian insufficiency as described in claim 1, characterized in that, The process of determining the number of predictive key points and the corresponding predictive attribute categories based on clinical data includes: Each predictive attribute category is determined according to the requirements of pregnancy prediction, and several predictive points are selected based on clinical data according to key points of ovarian reserve function, key points of treatment intervention, and key points related to pregnancy outcome. The hormone sensitivity index and follicle development index for each prediction point are determined based on several prediction points. The pregnancy potential at each prediction point is calculated based on the hormone sensitivity index and follicle development index, and the stability at each prediction point is also calculated. Key prediction points are selected from the prediction points based on the pregnancy potential and stability.

6. The method for establishing a model to predict the pregnancy rate of early-onset ovarian insufficiency as described in claim 5, characterized in that, Based on clinical data, several predictive points were selected for each category, including: The menstrual cycle is divided according to time series based on the patient's clinical data; Selecting key time points from the follicular phase, ovulation phase, and luteal phase of the menstrual cycle as crucial points for ovarian reserve function; The rate of change in hormone levels at each key point of ovarian reserve function was calculated using a hormone kinetic model based on data from key points of basic physiological characteristics.

7. The method for establishing a model to predict the pregnancy rate of early-onset ovarian insufficiency as described in claim 6, characterized in that, The dimensionality reduction improved feedforward neural network includes: The input layer is used to convert key points of basic physiological characteristics and key points of treatment intervention, such as age, AMH level, FSH value, treatment plan, hormone level data, hormone sensitivity index and follicle development index, into multidimensional data of prediction points. Dimensionality reduction layer is used to reduce the dimensionality of the multidimensional data of the prediction points to obtain the dimensionality-reduced data of the multidimensional data of the prediction points; Multiple hidden layers are used to extract abstract features through linear transformation; There are two output layers. One output layer is used to output the pregnancy prediction result based on the abstract features and the activation function. The other output layer is used to output the hormone level prediction result of the ovarian reserve function point based on the abstract features output by part of the hidden layer. The adjustment unit is used to adjust the bias matrix corresponding to the partially hidden layer based on the results calculated by the hormone kinetic model.

8. A model-building system for predicting pregnancy rates in patients with early-onset ovarian insufficiency, characterized in that, The system includes: The feature identification module is used to acquire clinical data of patients with early-onset ovarian insufficiency, and determine the number of predictive key points and the corresponding predictive attribute categories based on the clinical data. The predictive attribute categories include basic physiological characteristic key points, ovarian reserve function key points, treatment intervention key points, and pregnancy outcome related key points. The dimensionality reduction modeling module is used to establish a dimensionality reduction improved feedforward neural network when the total number of key points of basic physiological characteristics, key points of ovarian reserve function, key points of treatment intervention, and key points related to pregnancy outcome exceeds a first threshold. The dimensionality reduction improved feedforward neural network takes the data of the key points of basic physiological characteristics and key points of treatment intervention as input and outputs the predicted data of key points of ovarian reserve function and key points related to pregnancy outcome. The graph network construction module is used to build a GNN model when the sum of one-quarter of the number of basic physiological feature key points, ovarian reserve function key points, treatment intervention key points, and pregnancy outcome related key points is less than a first threshold. The GNN model establishes and updates the connection relationship between hierarchical nodes based on the hierarchical relationship of the cluster nodes of the basic physiological feature key points, ovarian reserve function key points, treatment intervention key points, and pregnancy outcome related key points. It continuously updates each hierarchical node based on the data of the basic physiological feature key points and obtains the predicted classification of ovarian reserve function key points and pregnancy outcome related key points based on the updated hierarchical nodes.

9. A device for establishing a model to predict the pregnancy rate in patients with early-onset ovarian insufficiency, characterized in that, The device includes: a memory, a processor, and a model building program for predicting the pregnancy rate of early-onset ovarian insufficiency, stored in the memory and executable on the processor, the model building program for predicting the pregnancy rate of early-onset ovarian insufficiency being configured to implement the steps of the model building method for predicting the pregnancy rate of early-onset ovarian insufficiency as described in any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a model-building program for predicting the pregnancy rate of early-onset ovarian insufficiency, which, when executed by a processor, implements the steps of the model-building method for predicting the pregnancy rate of early-onset ovarian insufficiency as described in any one of claims 1 to 7.

Images