Individualized ovulation promoting scheme intelligent decision support system oriented to assisted reproduction

Abstract

The invention relates to the technical field of artificial intelligence assisted medical treatment, and particularly discloses an assisted reproduction-oriented individualized ovulation promoting scheme intelligent decision support system, which comprises a patient multi-dimensional data integration module, an ovary-intima dual-system dynamic model, an individualized scheme optimization engine and a clinical interaction and execution feedback module. According to the method, ovarian response and intimal receptivity are predicted in a coupling mode, a Pareto optimal scheme set giving consideration to ovum obtaining targets and intimal synchronism is generated through a double-target optimization algorithm, quantitative decision support is provided for clinicians, and model self-adaptive updating is achieved by executing feedback. Model prediction and optimization suggestions are continuously close to the real reaction law of individual patients and evolution of group treatment outcome along with data accumulation, and a virtuous circle which is more intelligent is formed, so that the efficiency of an ovulation promoting cycle and the success rate of pregnancy are improved.

Description

Technical Field

[0001] This invention belongs to the field of artificial intelligence-assisted medical technology, specifically relating to an intelligent decision support system for individualized ovulation induction protocols for assisted reproduction. Background Technology

[0002] Assisted reproductive technology (ART) is an important medical field for solving infertility and fulfilling the desire to have children. Its core lies in obtaining and utilizing gametes and embryos through a series of medical interventions. Ovulation induction, a crucial step in this process, aims to obtain multiple mature oocytes through drug stimulation, laying the foundation for subsequent in-vitro fertilization and embryo transfer.

[0003] Developing individualized ovulation induction protocols is a core technological direction for improving the success rate of assisted reproductive technology. Its goal is to select and adjust the type, dosage, and timing of medication based on the patient's physiological characteristics, hormone levels, and ovarian responsiveness, in order to obtain the optimal number of high-quality eggs.

[0004] Decision support systems for ovulation induction protocols primarily rely on patient age, basal hormone levels, and antral follicle counts to predict ovarian responsiveness and recommend standard medication regimens. However, existing systems have significant limitations: their decision-making logic focuses almost entirely on assessing ovarian response to gonadotropins and the quantity and quality of retrieved eggs, generally neglecting the profound impact of ovulation-inducing drugs and their induced endocrine changes on endometrial receptivity—a crucial factor. Successful ovulation induction protocols depend not only on obtaining high-quality eggs but also on creating a synchronized and favorable endometrial environment for subsequent embryo implantation. Current systems treat ovulation induction and endometrial preparation as separate, independent stages, leading to situations where recommended protocols may yield good eggs but suffer from reduced overall pregnancy rates due to endometrial asynchrony or impaired receptivity. Summary of the Invention

[0005] The purpose of this invention is to provide an intelligent decision support system for individualized ovulation induction protocols for assisted reproduction, in order to solve the technical problem that existing ovulation induction protocol decision systems only focus on ovarian response and ignore its impact on endometrial receptivity, resulting in a disconnect between the ovulation induction and endometrial preparation stages, which may reduce the overall pregnancy rate.

[0006] This invention provides an intelligent decision support system for individualized ovulation induction protocols for assisted reproduction. This system is a computational decision platform integrating multi-source data acquisition, dynamic physiological modeling, dual-objective optimization, and risk warning functions. The system includes a patient multidimensional data integration module, an ovarian-endometrial dual-system dynamic model, an individualized protocol optimization engine, and a clinical interaction and execution feedback module.

[0007] The patient multidimensional data integration module is used to collect and structure patient data from different medical information systems in real time. This module specifically includes a basic profile unit, a cycle monitoring unit, and a historical cycle database. The basic profile unit is used to input and store the patient's static physiological parameters, including age, body mass index, infertility etiology classification, anti-Müllerian hormone concentration, basal antral follicle count, basal follicle-stimulating hormone concentration, basal luteinizing hormone concentration, and previous ovulation induction protocols and outcome records. The cycle monitoring unit is used to automatically or manually collect dynamic monitoring data at preset time points after the ovulation induction cycle begins. This dynamic monitoring data includes follicle diameter and number measured by transvaginal ultrasound, endometrial thickness and type, serum estradiol concentration, progesterone concentration, and luteinizing hormone concentration. The historical cycle database is used to archive and index complete datasets of all assisted reproductive cycles for the same patient, providing a data foundation for longitudinal comparative analysis.

[0008] The ovarian-endometrial dual-system dynamic model is the core computational unit of the system. It is used to construct and iteratively update a computational model representing the individual physiological state of a patient based on real-time and historical data provided by the patient's multidimensional data integration module. This model consists of a coupled sub-model of ovarian response prediction and a sub-model of endometrial receptivity prediction. The ovarian response prediction sub-model employs a time-series neural network trained on massive amounts of clinical cycle data. Its inputs are the follicle diameter, estradiol concentration, and medication records arranged in time series for the current cycle. Its outputs are the predicted follicle development trend, the expected range of oocyte retrieval, and the probability distribution of oocyte maturity over the next 24 to 72 hours. The endometrial receptivity prediction sub-model employs a fusion architecture of a pharmacokinetic-pharmacodynamic model based on physiological mechanisms and machine learning. Its inputs include the current cycle's medication regimen, time-series data of estradiol and progesterone concentrations, and changes in endometrial thickness and typing. By simulating the metabolic process of drugs in vivo and their regulatory effects on the gene expression profiles of endometrial epithelial cells and stromal cells, the outputs are the predicted endometrial receptivity window opening time and the comprehensive receptivity score within the window period. The comprehensive receptivity score is a weighted function value of endometrial thickness, blood perfusion index, histological score, and expression levels of specific biomarkers.

[0009] The individualized protocol optimization engine receives prediction results from the ovarian-endometrial dual-system dynamic model and performs real-time optimization calculations of the ovulation induction medication protocol based on the criteria of simultaneously optimizing oocyte retrieval and endometrial receptivity goals. This engine includes a dual-objective function construction unit, a constraint loading unit, and an optimization algorithm solution unit. The dual-objective function construction unit defines two core objectives for protocol optimization: the first objective function is to maximize the expected number of high-quality follicles, and the second objective function is to maximize the synchronicity between the predicted endometrial receptivity window opening date and the planned oocyte retrieval date, using the comprehensive receptivity score as a weighting factor for synchronicity.

[0010] The constraint loading unit loads inviolable hard constraints and recommended soft constraints. Hard constraints include the maximum tolerated dose of gonadotropins for patients, the risk threshold for ovarian hyperstimulation syndrome, and the drug use guidelines approved by the National Medical Products Administration that the protocol must comply with. Soft constraints include medication cost preferences and the memory of the effectiveness of previous protocols. The optimization algorithm solution unit employs a non-dominated sorting genetic algorithm with an elitist strategy. Within the solution space satisfying all hard constraints, it iteratively optimizes the bi-objective function over multiple generations, ultimately outputting a Pareto optimal solution set composed of multiple non-dominated solutions. Each solution corresponds to a complete set of candidate ovulation induction protocols, including drug type, starting dose, timing and magnitude of dose adjustments, and their corresponding bi-objective predicted values.

[0011] The clinical interaction and execution feedback module visualizes the Pareto optimal solution set output by the personalized protocol optimization engine, assisting clinicians in making final decisions and collecting real-world outcome data after decision execution for closed-loop system optimization. This module includes a protocol visualization decision interface, an execution tracking unit, and a model adaptive update unit. The protocol visualization decision interface displays the Pareto front in a two-dimensional coordinate system. The horizontal axis represents the expected number of high-quality follicles, and the vertical axis represents the endometrial receptivity synchronicity score. Each front point is associated with its detailed medication regimen and prediction details. Doctors can interactively select the protocol point that best suits their current clinical priorities and confirm execution. The execution tracking unit continuously compares actual monitoring data with model prediction data throughout the protocol execution cycle. When key indicators such as estradiol growth curves or endometrial thickness deviate from the predicted trajectory by more than a preset deviation threshold, the system is triggered to restart the personalized protocol optimization engine for mid-term protocol adjustment calculations. The model adaptive update unit is used to incrementally learn the neural network parameters and pharmacodynamic parameters in the ovarian-endometrial dual system dynamic model after the cycle ends, using the complete input data of this cycle, the final executed plan, and the actual outcome data, including the number of oocytes retrieved, oocyte maturity rate, fertilization rate, number of usable embryos, and the final pregnancy outcome, as new training samples. This enables the model to continuously evolve for individual patients and population trends.

[0012] In one embodiment of the present invention, the time-series neural network used in the ovarian response prediction sub-model has a multi-layer encoder-decoder structure comprising a long short-term memory network layer and an attention mechanism. The encoder is used to extract features and encode context from the input time-series monitoring data. The long short-term memory network layer is responsible for capturing the long-term dependencies between follicle development and hormonal changes. The attention mechanism dynamically assigns different weights to features at different historical time points when decoding and predicting future states, focusing on recent trends and key turning points. The decoder, based on the encoded context vector, progressively generates prediction sequences for multiple future time points, outputting follicle diameter distribution, predicted estradiol concentration, and the probability of occurrence of follicle maturation marker events.

[0013] Furthermore, the pharmacokinetic-pharmacodynamic model within the endometrial receptivity prediction sub-model specifically simulates the absorption, distribution, metabolism, and excretion of gonadotropin-releasing hormone agonists or antagonists, urinary or recombinant gonadotropins, and human chorionic gonadotropin in the patient's body. The pharmacokinetic component employs individualized compartmental model parameters based on patient body mass index and renal function estimations. The pharmacodynamic component maps the calculated plasma drug concentration time-course data to the stimulatory effect on estradiol synthase activity in ovarian granulosa cells and the intensity of its influence on progesterone receptor expression and pinocytosis-related pathways in endometrial epithelial cells. This mapping relationship was determined through regression analysis integrating publicly available molecular biology databases and historical data from our hospital.

[0014] Furthermore, the dual-objective function construction unit in the individualized treatment optimization engine has a second objective function mathematically expressed in terms of synchronicity as follows: It calculates the absolute value of the number of days between the predicted endometrial receptivity window opening date and the planned egg retrieval date, performs a negative exponential transformation on this absolute value, and then multiplies it by the comprehensive receptivity score on the window opening date. This is used as the synchronicity score. This mathematical expression means that the closer the window opening date is to the egg retrieval date, and the higher the endometrial quality on that day, the higher the synchronicity score.

[0015] Furthermore, the execution tracking unit within the clinical interaction and execution feedback module employs a dynamic adaptive method for setting its deviation threshold. This method is determined based on the number of data points monitored in the current cycle and the historical confidence interval of the model's prediction at that time point. Specifically, as the cycle progresses, the accumulated data points increase, the confidence interval of the model's prediction gradually narrows, and the corresponding deviation threshold also decreases synchronously. This makes the system more sensitive to prediction deviations in the mid-to-late stages of the cycle than in the early stages, thereby more accurately capturing clinically significant deviations and triggering timely protocol re-optimization.

[0016] Furthermore, the system is deployed on a cloud server. The patient multidimensional data integration module securely interfaces with the hospital's laboratory information system, image archiving and communication system, and electronic medical record system through an application programming interface that conforms to medical and health information exchange standards. The clinical interaction and execution feedback module's visual decision-making interface is accessible to authorized physicians in web page or mobile application format, and all data transmission is encrypted.

[0017] Compared with the prior art, the beneficial effects of the present invention are as follows: 1. This invention constructs an intelligent decision support system that deeply integrates the dual physiological systems of ovarian response and endometrial receptivity, fundamentally changing the limitations of the traditional separation of these two systems in ovulation induction protocol formulation. By establishing a dynamic model of the ovarian-endometrial dual system, the system can quantitatively predict the synergistic evolution of follicular development and endometrial receptivity window under different medication regimens, achieving a more comprehensive physiological simulation of the entire ovulation induction process.

[0018] 2. The personalized protocol optimization engine adopts a multi-objective optimization algorithm, which can actively explore a set of Pareto optimal protocols that can simultaneously take into account the quantity and quality of retrieved eggs and the synchronicity of endometrial preparation, while meeting clinical safety constraints. This provides doctors with a scientific decision-making basis based on quantitative trade-offs, rather than a recommendation based on a single objective.

[0019] 3. The execution tracking and model adaptive update mechanism ensures that the system can continuously learn from actual clinical practice, so that the model predictions and optimization suggestions become closer to the real response patterns of individual patients and the evolution of group treatment outcomes as data accumulates, forming a virtuous cycle of becoming smarter with use.

[0020] 4. Elevating the ovulation induction process in assisted reproduction from an experience-dependent and relatively singular decision-making process to a data-driven, globally optimized, and self-evolving precision medicine practice is expected to significantly improve the efficiency of ovulation induction cycles and the final pregnancy success rate. Attached Figure Description

[0021] Figure 1 This is a schematic diagram of the overall technical solution architecture of the present invention; Figure 2 This is a schematic diagram of the core principle framework of the dynamic model of the ovarian-endometrial dual system in this invention; Figure 3 This is a flowchart of the dual-objective optimization logic of the individualized scheme optimization engine in this invention. Figure 4 This is a schematic diagram of the closed-loop interaction and data flow of the clinical interaction and execution feedback module in this invention; Figure 5 This is a schematic diagram illustrating the dynamic adaptive deviation threshold setting principle of the system execution tracking unit in this invention. Detailed Implementation

[0022] Example 1: The overall architecture of the intelligent decision support system for individualized ovulation induction protocols for assisted reproduction proposed in this invention is shown in the attached figure. Figure 1 As shown, this is a computational decision-making platform integrating multi-source data acquisition, dynamic physiological modeling, dual-objective optimization, and risk warning functions. The system constructs a complete technology chain from data input, physiological simulation, and protocol generation to clinical execution and feedback closed-loop through four core components: a patient multi-dimensional data integration module, an ovarian-endometrial dual-system dynamic model, a personalized protocol optimization engine, and a clinical interaction and execution feedback module. The following will combine the attached... Figures 1 to 5 The specific implementation methods of each component of the system are described in detail.

[0023] The patient multidimensional data integration module is the data foundation for the entire system's operation, and its structure is shown in the attached figure. Figure 1 As shown, the system consists of three sub-units: a basic profile unit, a cycle monitoring unit, and a historical cycle database. The basic profile unit is used to input and store the patient's static physiological parameters. These parameters are collected and stored in the system database when the patient first enters the assisted reproductive treatment process. These static physiological parameters include, but are not limited to: age, body mass index (BMI), infertility etiology diagnosis classification (standardized using the International Classification of Diseases coding system), anti-Müllerian hormone (AMH) concentration, basal antral follicle count, basal follicle-stimulating hormone (FSH) concentration (in ISUs per liter), basal luteinizing hormone (LH) concentration, and previous ovulation induction protocols and outcome records (including structured fields such as drug type, initial dose, adjustment strategy, number of oocytes retrieved, embryo development status, and pregnancy outcome). All static parameters must undergo a data verification mechanism to confirm their reasonableness; for example, a BMI exceeding the range of 15 to 40 will trigger a manual review process.

[0024] The cycle monitoring unit is responsible for automatically or manually collecting dynamic monitoring data at preset time points after the start of the current ovulation induction cycle. This unit establishes a secure data interface with the hospital's laboratory information system and image archiving and communication system to obtain key physiological indicators in real time. The dynamic monitoring data includes: follicle development diameter and number measured by transvaginal ultrasound (grouped by diameter, such as ≥10 mm, ≥14 mm, ≥18 mm, etc.), endometrial thickness and type (classified as type A, B, and C based on the three-line sign), serum estradiol concentration, progesterone concentration, and luteinizing hormone concentration. All dynamic data are accurately timestamped and stored in time-series format to ensure that subsequent models can accurately capture the evolution of physiological states. The data collection frequency is dynamically adjusted according to the ovulation induction stage, once daily in the early follicular phase and increased to once every 12 hours as the dominant follicle selection period approaches.

[0025] The historical cycle database archives and indexes complete datasets of all assisted reproductive cycles for the same patient, forming a longitudinal patient health record. Each historical cycle record contains complete static parameters, dynamic monitoring sequences, execution protocol details, and final clinical outcomes. The system achieves cross-cycle data association through unique patient identifiers and supports rapid retrieval and comparative analysis by multiple dimensions such as treatment year, protocol type, and outcome category. The historical cycle database not only provides a reference benchmark for the current cycle but also serves as an important source of training data for adaptive model updates.

[0026] The ovarian-endometrial dual-system dynamic model is the core computational unit of this system. Its internal structure and data flow are shown in the attached figure. Figure 2 As shown, the model consists of a coupled sub-model for predicting ovarian response and a sub-model for predicting endometrial receptivity. These two sub-models share some input data but are modeled separately, ultimately outputting a collaborative prediction result. The ovarian response prediction sub-model uses a time-series neural network trained on massive amounts of clinical cycle data. Its specific architecture is a multi-layer encoder-decoder structure containing a long short-term memory network layer and an attention mechanism. The encoder receives follicle diameter, estradiol concentration, and medication records arranged in time series for the current cycle as input. The long short-term memory network layer effectively captures the long-term dependency between follicle development speed and hormone level changes through a gating mechanism, avoiding the gradient vanishing problem in traditional recurrent neural networks. The attention mechanism dynamically calculates the importance weights of features at each historical time point during the decoding stage, enabling the model to focus on recent key turning points (such as a sharp rise in estradiol or the establishment of a dominant follicle) rather than treating all historical data uniformly when predicting future states.

[0027] The decoder, based on the context vector output by the encoder, progressively generates predicted sequences for multiple time points within the next 24 to 72 hours. The output includes: the distribution of follicle diameters at each future time point (represented as a probability density function), predicted estradiol concentrations (with a 95% confidence interval), and the probability of occurrence of follicle maturation marker events (e.g., diameter ≥ 18 mm and estradiol > 917 picomoles per liter). The training data for this sub-model comes from high-quality labeled data of over 100,000 completed cycles in a historical cycle database. During training, a weighted combination of mean squared error and cross-entropy loss is used to ensure a balance in prediction accuracy for continuous variables and categorical events.

[0028] The endometrial receptivity prediction sub-model employs a fusion architecture of pharmacokinetic-pharmacodynamic model and machine learning based on physiological mechanisms. Its inputs include the current cycle's medication regimen (drug type, dosage, route of administration, and timing of administration), time-series concentration data of estradiol and progesterone, and changes in endometrial thickness and typing. The pharmacokinetic component first simulates the absorption, distribution, metabolism, and excretion of gonadotropin-releasing hormone agonists or antagonists, urinary or recombinant gonadotropins, and human chorionic gonadotropin (hCG) in the patient's body. This process uses individualized compartmental model parameters determined based on the patient's body mass index and estimated glomerular filtration rate to convert the dosing regimen into time-series plasma drug concentration curves.

[0029] The pharmacodynamics section maps the calculated blood drug concentration time-course data to the stimulatory effect on estradiol synthase activity in ovarian granulosa cells and the intensity of its influence on progesterone receptor expression and pinocytosis-related pathways in endometrial epithelial cells. This mapping relationship is determined through multivariate regression analysis integrating publicly available molecular biology databases (such as KEGG and Reactome) and historical data from our hospital, forming an interpretable physiological response function. Based on this, the model further integrates endometrial thickness, blood perfusion index (obtained via Doppler ultrasound), histological morphology score (assessed by a senior reproductive ultrasound physician according to standard atlases), and expression levels of specific biomarkers (such as integrin αvβ3 and leukemia inhibitory factor), calculating a comprehensive receptivity score using a weighted function. This score ranges from 0 to 100, with higher scores indicating the endometrium is closer to the ideal implantation state. The model ultimately outputs the predicted endometrial receptivity window opening time (usually defined as the time when the comprehensive receptivity score first consistently exceeds 75 points) and the daily comprehensive receptivity score within the window period.

[0030] The personalized protocol optimization engine receives prediction results from the ovarian-endometrial dual-system dynamic model and performs real-time optimization calculations of the ovulation induction medication protocol based on the criteria of simultaneously optimizing oocyte retrieval and endometrial receptivity goals. Its internal logic is shown in the attached figure. Figure 3 As shown, the algorithm consists of three parts: a dual-objective function construction unit, a constraint loading unit, and an optimization algorithm solution unit. The dual-objective function construction unit defines two core objectives for the optimization scheme: the first objective function is to maximize the expected number of high-quality follicles, which is calculated by the probability-weighted average of the number of follicles with a diameter ≥14 mm output by the ovarian response prediction sub-model; the second objective function is to maximize the synchronicity between the predicted endometrial receptivity window opening date and the planned oocyte retrieval date. The specific mathematical expression for synchronicity is as follows: in, To score the synchronization, The predicted opening date of the endometrial receptivity window. For the planned egg retrieval date, The overall tolerance score for the day the window opens. The attenuation coefficient, set to 0.5, is used to adjust the penalty intensity of time deviation on the score. This formula indicates that the closer the window opening date is to the egg retrieval date, and the higher the endometrial quality on that day, the higher the synchronization score.

[0031] The constraint loading unit is responsible for loading inviolable hard constraints and recommended soft constraints. Hard constraints include: the maximum tolerated dose of gonadotropins for patients (determined by referring to a table based on age and anti-Müllerian hormone concentration, typically 150 to 450 IU daily), the risk threshold for ovarian hyperstimulation syndrome (a high-risk warning is triggered when the predicted estradiol concentration is >15,000 picoseconds per liter or the number of follicles is >20), and the requirement that the protocol must comply with the drug use guidelines approved by the National Medical Products Administration (e.g., gonadotropin-releasing hormone antagonists can only be used after day 5 of the follicular phase). Soft constraints include medication cost preferences (users can set a budget cap, and the system prioritizes protocols with costs below that cap) and the memory of the effectiveness of previous protocols (if a protocol in a patient's past cycles has achieved good outcomes, the parameters of that protocol will be given a higher initial weight in the current cycle optimization).

[0032] The optimization algorithm employs a non-dominated sorting genetic algorithm with an elitist strategy. During algorithm initialization, 200 candidate schemes are randomly generated within the solution space satisfying all hard constraints as the initial population. Each scheme is encoded as a vector containing the drug type, initial dose, and timing and magnitude of dose adjustment. The algorithm iterates for 50 generations, dividing the population into multiple frontier levels each generation through non-dominated sorting and maintaining solution diversity using crowding distance. The elitist strategy ensures that the best 20% of individuals in each generation directly enter the next generation, avoiding the loss of high-quality solutions. The final output is a Pareto optimal solution set consisting of 10 to 15 non-dominated solutions, each solution corresponding to a complete set of candidate ovulation induction protocols and its corresponding dual-objective predicted values ​​(expected number of high-quality follicles and synchronicity score).

[0033] The Clinical Interaction and Execution Feedback module is responsible for transforming optimization results into clinically usable decision support information and constructing a closed-loop feedback mechanism. Its overall interaction flow is shown in the attached figure. Figure 4As shown, the protocol visualization decision interface displays the Pareto front in a two-dimensional coordinate system. The horizontal axis represents the expected number of high-quality follicles, and the vertical axis represents the endometrial receptivity synchronicity score (dimensionless, ranging from 0 to 1). Each front point is represented by a circular icon. Hovering the mouse over it displays its detailed medication regimen (e.g., recombinant gonadotropin 225 IU daily, with an antagonist added on day 6) and prediction details (e.g., expected oocyte retrieval of 12±2, window opening one day before oocyte retrieval, receptivity score of 82). Doctors can select the protocol point that best suits their current clinical priority (e.g., prioritizing protocols with high synchronicity in older patients, and prioritizing protocols with a higher number of oocytes retrieved in younger, low-response patients) and confirm execution. The system then generates structured medical orders and pushes them to the hospital's electronic medical record system.

[0034] The execution tracking unit continuously compares actual monitoring data with model prediction data throughout the program's execution cycle. (See attached...) Figure 5 As shown, the deviation threshold is set using a dynamic adaptive method. This method is determined based on the number of monitored data points in the current cycle and the historical confidence interval of the model prediction at that time point. Specifically, the system maintains a historical prediction error database, recording the standard deviation of the difference between the predicted and actual values ​​for similar patients on the same cycle day. The deviation threshold for day n of the current cycle... The calculation formula is: in, This represents the standard deviation of the prediction error for similar patients on day n in the historical data. This represents the total number of days in a typical ovulation induction cycle (usually 12 days). The sensitivity coefficient is set to 1.5. This formula indicates that as the cycle progresses (n increases), the deviation threshold gradually decreases, making the system more sensitive to prediction deviations in the mid-to-late stages of the cycle than in the early stages. When key indicators such as estradiol growth curves or endometrial thickness deviate from the predicted trajectory by more than [a certain value], [the system becomes more sensitive to these deviations]. When this happens, the system automatically triggers an alarm and restarts the individualized treatment optimization engine to perform mid-term treatment adjustment calculations and generate a new Pareto frontier for physicians to make decisions.

[0035] The adaptive update unit of the model comes into play after the cycle ends. The system automatically collects complete input data, the final executed protocol, and actual outcome data for this cycle, including the number of oocytes retrieved, oocyte maturation rate (number of mature oocytes / total number of oocytes retrieved), fertilization rate (number of normally fertilized oocytes / number of mature oocytes), number of usable embryos (number of embryos that have developed to blastocysts and have a morphological score ≥3BB), and the final pregnancy outcome (biochemical pregnancy, clinical pregnancy, live birth). This complete dataset is packaged into new training samples for incremental learning of the ovarian-endometrial dual-system dynamic model. For the ovarian response prediction sub-model, the weight parameters of the long short-term memory network layer are fine-tuned using online gradient descent; for the endometrial receptivity prediction sub-model, the regression coefficients in the pharmacodynamic mapping function are adjusted through Bayesian updates. This process ensures that the model can continuously adapt to the unique physiological response patterns of individual patients and the evolution of group treatment trends.

[0036] The entire system is deployed on a cloud server and employs a microservice architecture to achieve loose coupling between modules. The patient multidimensional data integration module securely interfaces with the hospital's laboratory information system, image archiving and communication system, and electronic medical record system via an application programming interface (API) compliant with healthcare information exchange standards (such as HL7 FHIR). All interfaces utilize two-way SSL certificate authentication and AES-256 encrypted transmission. The clinical interaction and execution feedback module's visual decision-making interface is developed as a responsive web application, compatible with mainstream browsers, and also provides a mobile application version for authorized physicians to access in the clinic or remotely. The system has a strict access control mechanism to ensure that patient data is only accessible to researchers approved by the ethics committee and the medical team directly responsible for treatment.

[0037] Through the collaborative work of the aforementioned modules, this system achieves a paradigm shift from solely focusing on ovarian response to simultaneously optimizing the ovarian-endometrial dual system. In practical clinical applications, the system not only provides quantitative and visualized criteria for treatment selection but also transforms each treatment cycle into valuable data for enhancing model intelligence through execution tracking and adaptive update mechanisms, ultimately forming an intelligent decision support ecosystem that becomes increasingly precise and personalized with each use.

[0038] Example 2: Building upon Example 1, this example further refines the parameter calibration mechanism of the pharmacokinetic-pharmacodynamic model in the endometrial receptivity prediction sub-model and introduces a multimodal biomarker fusion strategy to improve prediction robustness. Specifically, the compartmental model parameters in the pharmacokinetic part no longer rely solely on body mass index and renal function estimation, but are finely adjusted in conjunction with the patient's liver enzyme gene polymorphism detection results. The system adds a drug metabolism genotype field to the basic file unit to store the detection results of single nucleotide polymorphism sites of key metabolic enzymes such as CYP3A4 and CYP3A5. When a patient provides such a gene testing report, the system dynamically corrects the drug clearance rate and volume of distribution parameters based on the established genotype-pharmacokinetic parameter association table. For example, for patients carrying CYP3A5... 3 / In patients with homozygous mutations, the clearance rate of gonadotropin-releasing hormone antagonists will be reduced by 30%, resulting in a prolonged period of blood drug concentration maintenance, which in turn affects the prediction of the endometrial receptivity window.

[0039] In the pharmacodynamic mapping phase, this embodiment introduces a multimodal biomarker fusion strategy. In addition to traditional ultrasound and hormone indicators, the system supports the integration of endometrial liquid-based cytology data. This test involves collecting uterine lavage fluid on a specific day of the menstrual cycle and quantitatively detecting leukemia inhibitory factors and integrins using mass spectrometry analysis. 3. Concentrations of 12 receptivity-related proteins, including osteopontin. These protein concentration data, after standardization, were fed as additional input features into the machine learning branch of the endometrial receptivity prediction sub-model. This branch employs a gradient boosting tree model to non-linearly fuse ultrasound, hormone, and protein biomarker data, outputting a corrected comprehensive receptivity score. The fusion weights are automatically learned on historical data through cross-validation to ensure optimal predictive performance across different patient subgroups.

[0040] Furthermore, the soft constraint loading mechanism in the personalized treatment optimization engine has been enhanced. A new patient preference sub-unit has been added, allowing patients to express their subjective preferences regarding the treatment process through a dedicated questionnaire before treatment, such as minimizing the number of injections or accepting slightly higher costs for a higher success rate. These preferences are quantified as additional weighting factors for the optimization objectives. For example, if a patient chooses to reduce the number of injections, during the optimization process, regimens containing long-acting preparations (such as pegylated gonadotropins) are given higher priority, even if their costs are slightly higher. This mechanism ensures that system decisions are based not only on objective physiological data but also incorporate a patient-centered treatment philosophy.

[0041] The deviation detection logic of the execution tracking unit has also been upgraded. In addition to single-indicator deviation detection, a multi-indicator collaborative deviation detection module has been added to the system. This module uses principal component analysis to project multiple dynamic indicators such as estradiol, progesterone, endometrial thickness, and follicle count into a low-dimensional feature space, and calculates the Mahalanobis distance between the current state point and the model's predicted trajectory in this space. When this distance exceeds a critical value determined based on the chi-square distribution, even if each individual indicator does not exceed the threshold, the system still determines that there is a potential risk of physiological disorder and triggers mid-term protocol re-optimization. This strategy effectively improves the ability to identify complex and atypical response patterns.

[0042] Through the above enhancements, this embodiment significantly improves the system's adaptability and prediction accuracy in scenarios with diverse genetic backgrounds and abundant biomarker data, while maintaining the core architecture unchanged.

Claims

1. An individualized ovulation induction protocol smart decision support system for assisted reproduction, characterized in that, include: The patient multidimensional data integration module is used to collect and structure patient real-time and historical data from different medical information systems. The ovarian-endometrial dual-system dynamic model is used to construct and iteratively update a computational model characterizing the individual physiological state of a patient based on real-time and historical data provided by the patient multidimensional data integration module. The individualized protocol optimization engine receives the results output by the dynamic model of the ovarian-endometrial dual system, and performs real-time optimization calculations of the ovulation induction drug protocol based on the criteria of simultaneously optimizing the oocyte retrieval target and the endometrial receptivity target, and outputs the Pareto optimal solution set; The clinical interaction and execution feedback module is used to visualize the Pareto optimal solution set output by the individualized solution optimization engine, assist clinicians in making final decisions, and collect real-world outcome data after decision execution to optimize the system in a closed loop.

2. The individualized ovulation induction protocol for assisted reproduction approach intelligent decision support system according to claim 1, characterized in that, The patient multidimensional data integration module includes a basic file unit, a periodic monitoring unit, and a historical period database. The basic file unit is used to input and store the patient's static physiological parameters, which include age, body mass index, infertility etiology diagnosis classification, anti-Müllerian hormone concentration, basal antral follicle count, basal follicle-stimulating hormone concentration, basal luteinizing hormone concentration, and previous ovulation induction protocols and outcome records. The cycle monitoring unit is used to collect dynamic monitoring data at preset time nodes after the ovulation induction cycle is started. The dynamic monitoring data includes the diameter and number of follicles, endometrial thickness and type, serum estradiol concentration, progesterone concentration and luteinizing hormone concentration measured by transvaginal ultrasound. The historical cycle database is used to archive and index the complete dataset of all assisted reproductive cycles of the same patient.

3. The individualized ovulation induction protocol for assisted reproduction approach intelligent decision support system according to claim 2, characterized in that, The dynamic model of the ovarian-endometrial dual system is composed of a coupled sub-model of ovarian response prediction and a sub-model of endometrial receptivity prediction. The sub-model of ovarian response prediction uses a time-series neural network trained with massive clinical cycle data. Its input is the follicle diameter, estradiol concentration and medication records of the current cycle arranged in time series. Its output is the predicted value of future follicle development trend, the range of expected number of oocytes retrieved and the probability distribution of oocyte maturity.

4. The individualized ovulation induction protocol for assisted reproduction approach intelligent decision support system according to claim 3, characterized in that, The endometrial receptivity prediction sub-model adopts a fusion architecture of pharmacokinetic-pharmacodynamic model based on physiological mechanisms and machine learning. Its inputs are the current cycle's medication regimen, the time-series data of estradiol and progesterone concentrations, and changes in endometrial thickness and classification. The outputs are the predicted endometrial receptivity window opening time and the comprehensive receptivity score during the window period.

5. The individualized ovulation induction protocol for assisted reproduction purpose smart decision support system according to claim 4, characterized in that, The individualized scheme optimization engine includes a dual objective function construction unit, a constraint loading unit, and an optimization algorithm solution unit; The dual-objective function construction unit is used to define two core objectives for scheme optimization. The first objective function is to maximize the expected number of high-quality follicles, and the second objective function is to maximize the synchronicity between the predicted endometrial receptivity window opening date and the planned egg retrieval date, and the comprehensive receptivity score is used as a weighting factor for synchronicity. The constraint loading unit is used to load inviolable hard constraints and recommended soft constraints; the optimization algorithm solving unit adopts a non-dominated sorting genetic algorithm with an elite strategy to perform multi-generation iteration optimization of the bi-objective function in the solution space that satisfies all hard constraints, and finally outputs a Pareto optimal solution set composed of multiple non-dominated solutions.

6. The individualized ovulation induction protocol for assisted reproduction purposes smart decision support system according to claim 5, characterized in that, The clinical interaction and execution feedback module includes a protocol visualization decision interface, an execution tracking unit, and a model adaptive update unit. The visualization decision interface of the proposed scheme displays the Pareto front in the form of a two-dimensional coordinate system, with the horizontal axis representing the expected number of high-quality follicles and the vertical axis representing the endometrial receptivity synchronicity score. The execution tracking unit is used to continuously compare the actual monitoring data with the model's predicted data during the scheme execution cycle. When the key indicators deviate from the predicted trajectory by more than a preset deviation threshold, the system is triggered to restart the individualized scheme optimization engine to perform mid-term scheme adjustment calculations. The model adaptive update unit is used to incrementally learn the parameters in the ovarian-endometrial dual-system dynamic model by using the complete input data of this cycle, the final executed plan, and the actual outcome data as new training samples after the cycle ends.

7. The individualized ovulation induction protocol for assisted reproduction purposes smart decision support system according to claim 6, characterized in that, The time-series neural network used in the ovarian response prediction sub-model has an architecture that includes a multi-layer encoder-decoder structure containing a long short-term memory network layer and an attention mechanism. The encoder section is used to extract features and encode contextual information from the input time-series monitoring data; the long short-term memory network layer is responsible for capturing the long-term dependencies between follicle development and hormonal changes. The attention mechanism is used to dynamically assign different weights to features at different historical time points when decoding and predicting future states; The decoder part generates prediction sequences for multiple future time points step by step based on the encoded context vector.

8. The individualized ovulation induction protocol for assisted reproduction approach intelligent decision support system according to claim 7, characterized in that, The pharmacokinetic-pharmacodynamic model in the endometrial receptivity prediction sub-model specifically simulates the absorption, distribution, metabolism, and excretion of gonadotropin-releasing hormone agonists or antagonists, urinary or recombinant gonadotropins, and human chorionic gonadotropin in the patient's body. The pharmacokinetic section uses individualized compartment model parameters based on patient body mass index and renal function estimation; the pharmacodynamic section maps the calculated blood drug concentration time-course data to the stimulatory effect on estradiol synthase activity in ovarian granulosa cells, and the intensity of the effect on progesterone receptor expression and pinocytosis-related pathways in endometrial epithelial cells.

9. The intelligent decision support system for individualized ovulation induction protocols for assisted reproduction according to claim 8, characterized in that, In the dual-objective function construction unit, the specific mathematical expression of the second objective function with respect to synchronicity is as follows: calculate the absolute value of the number of days between the predicted endometrial receptivity window opening date and the planned egg retrieval date, and then multiply this absolute value by the comprehensive receptivity score on the window opening date to obtain the synchronicity score.

10. The intelligent decision support system for individualized ovulation induction protocols for assisted reproduction according to claim 9, characterized in that, In the constraint loading unit, the hard constraints include the maximum tolerated dose of gonadotropins by the patient, the risk threshold of ovarian hyperstimulation syndrome, and the drug use guidelines approved by the National Medical Products Administration that the regimen must comply with; the soft constraints include medication cost preferences and memory of the effectiveness of previous regimens.

Images