Individualized ovulation promoting scheme and gonadotropin initial dose prediction method based on multi-cycle time sequence modeling and related equipment

Abstract

The invention provides an individualized ovulation promoting scheme and gonadotropin initial dose prediction method based on multi-cycle time sequence modeling and related equipment, and is applied to the technical field of data processing. The method comprises the following steps: firstly, receiving multi-cycle clinical data and basic information of a target user for IVF / ICSI pregnancy-assisted patients, and eliminating unqualified samples to generate an initial data set; after SHAP screening features, training / test set division and SMOTE sample equalization are carried out, an initial model is constructed through a sliding window time sequence module and a multi-algorithm engine; a target model is generated through training and evaluation, and a COS scheme prediction signal and multi-period Gn dose adjustment information are generated in combination with hierarchical classification and weighted integration; calibration information is generated through clinical data verification, after target user information is subjected to standardization processing, an individualized COS scheme and a Gn initial dose are output through a target model, and multi-cycle dynamic accurate recommendation is achieved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of data processing, in particular to an individualized ovulation promotion scheme and prediction method of starting dose of gonadotropin based on multi-cycle time sequence modeling and related equipment. BACKGROUND

[0002] About 1 / 6 of the global population of childbearing age is affected by infertility, which has become a major public health challenge. In recent years, influenced by factors such as changes in population structure, environmental pollution, and changes in lifestyle, the incidence of infertility has continued to rise, and coupled with the increasing awareness of reproductive health, patients' demand for assisted reproductive technology (ART) has grown. As a core means of ART, in vitro fertilization (IVF), the key link is controlled ovarian stimulation (COS), that is, through exogenous gonadotropin (Gn) stimulation to develop multiple follicles to obtain sufficient oocytes and improve the success rate of embryo transfer.

[0003] The current mainstream COS scheme includes GnRH-a scheme, GnRH-A scheme, and micro-stimulation scheme, and the starting dose of Gn directly affects the treatment effect. Due to significant individual differences among patients, improper doses can easily induce ovarian hyperstimulation syndrome (OHSS) or lead to ovarian hyporesponsiveness, so individualized COS scheme and dose formulation are the key to optimizing IVF effect.

[0004] Existing guidelines and expert consensus are difficult to fully cover complex individual differences, and the recommended opinions are scattered, with limited clinical practicability. Although there are studies attempting to build a Gn dose prediction model through machine learning, there are obvious limitations: the sample size is mostly several hundred cases, and the generalization ability is insufficient; it relies on linear regression models, which are difficult to analyze the non-linear relationship between Gn and ovarian response and the multi-factor interaction; it only focuses on the prediction of the first cycle Gn dose, lacks a COS scheme individualization selection model, and lacks multi-cycle dynamic guidance; the model has poor extrapolation, and the stability is insufficient in different clinical environments, with low actual application rate.

[0005] It should be noted that the information disclosed in the above background section is only used to strengthen the understanding of the background of the present disclosure, and therefore can include information that does not constitute prior art known to those of ordinary skill in the art. SUMMARY

[0006] Other characteristics and advantages of the present application will become apparent from the following detailed description, or will be learned by practice of the application.

[0007] According to an aspect of the present application, a method for predicting individualized ovulation induction scheme and gonadotropin starting dose based on multi-cycle time series modeling is provided, comprising: receiving multi-cycle clinical data of IVF / ICSI assisted reproduction treatment patients and target user basic information, while excluding patients with reproductive organ malformation, reproductive system tumor or chromosomal abnormality, generating an initial patient data set; using SHAP method to screen key predictors and quantify feature contribution, calling multi-cycle clinical data in the initial patient data set, dividing training set and test set according to preset processing rules, receiving pre-stored sample balance optimization algorithm and model evaluation rules, processing the training set to generate balanced training samples, starting sliding window time series processing module and multi-algorithm model training engine, generating patient historical cycle feature correlation data and initial individualized ovulation induction strategy recommendation model; training the initial individualized ovulation induction strategy recommendation model using the balanced training samples and the test set, combining the model evaluation rules to output the training results, generating a target ovulation prediction model, wherein the target ovulation prediction model includes an individualized ovulation induction strategy recommendation model and a Gn starting dose prediction model; processing the balanced training samples, test set evaluation results based on hierarchical classification strategy and weighted integration rules, generating COS scheme classification prediction signal and multi-cycle Gn dose prediction adjustment information, wherein the Gn starting dose prediction model is independently modeled according to the first to fourth and subsequent cycles; processing the COS scheme classification prediction signal, multi-cycle Gn dose prediction adjustment information and clinical applicability verification data, generating individualized scheme matching degree score and dose recommendation precision calibration information; cleaning and standardizing the target user basic information to generate user input data for prediction; processing the user input data for prediction, individualized scheme matching degree score and dose recommendation precision calibration information based on the target ovulation prediction model, generating and outputting individualized COS scheme and Gn starting dose recommendation results.

[0008] In another aspect of the present application, a device for predicting an individualized ovulation induction scheme and a starting dose of gonadotropin based on multi-cycle time series modeling comprises: an acquisition module for receiving multi-cycle clinical data of IVF / ICSI assisted reproduction treatment patients and target user basic information, while excluding patient samples with reproductive organ malformation, reproductive system tumor or chromosomal abnormalities, to generate an initial patient data set; a processing module for screening key prediction factors and quantifying feature contribution using the SHAP method, calling multi-cycle clinical data in the initial patient data set, dividing a training set and a test set according to a preset processing rule, receiving a pre-stored sample balancing optimization algorithm and model evaluation rule, processing the training set to generate a balanced training sample, starting a sliding window time series processing module and a multi-algorithm model training engine, and generating patient historical cycle feature correlation data and an initial individualized ovulation induction strategy recommendation model; training the initial individualized ovulation induction strategy recommendation model using the balanced training sample and the test set, outputting the training result in combination with the model evaluation rule to generate a target ovulation prediction model; processing the balanced training sample and the test set evaluation result based on a hierarchical classification strategy and a weighted integration rule to generate a COS scheme classification prediction signal and multi-cycle Gn dose prediction adjustment information; processing the COS scheme classification prediction signal, the multi-cycle Gn dose prediction adjustment information and the clinical applicability verification data to generate an individualized scheme matching degree score and dose recommendation precision calibration information; cleaning and standardizing the target user basic information to generate user input data that can be used for prediction; processing the user input data that can be used for prediction and the individualized scheme matching degree score and dose recommendation precision calibration information based on the target ovulation prediction model to generate and output an individualized COS scheme and a Gn starting dose recommendation result.

[0009] According to still another aspect of the present application, an electronic device comprises: a first processor; and a memory for storing executable instructions of the first processor; wherein the first processor is configured to execute the executable instructions to implement the above-mentioned prediction method of an individualized ovulation induction scheme and a starting dose of gonadotropin based on multi-cycle time series modeling.

[0010] According to still another aspect of the present application, an electronic device comprises: a first processor; and a memory for storing executable instructions of the first processor; wherein the first processor is configured to execute the executable instructions to implement the above-mentioned prediction method of an individualized ovulation induction scheme and a starting dose of gonadotropin based on multi-cycle time series modeling.

[0011] The application provides an individualized ovulation induction scheme based on multi-cycle time sequence modeling and a prediction method for gonadotropin starting dose and related equipment. The application introduces a comprehensive COS treatment personalized decision-making framework, constructs an individualized ovulation induction strategy recommendation model and a Gn starting dose prediction model, so that the scheme selection and the subsequent Gn starting dose can be individually customized, and the research gap that there is no prediction model for individualized selection of COS schemes at present is made up. In the development process of the prediction model, the application comprehensively uses various advanced machine learning algorithms, which can more effectively analyze the complex nonlinear relationship between the Gn starting dose and the number of oocytes, so that the constructed prediction model has higher accuracy and reliability. The ovulation induction scheme selection model based on the sliding window sequence algorithm has an accuracy of 76%; the multi-cycle Gn dose prediction model based on the LightGBM algorithm has a R²=0.79, and has good prediction performance and strong clinical applicability.

[0012] It should be understood that the above general description and the following detailed description are only exemplary and explanatory, and cannot limit the present disclosure. BRIEF DESCRIPTION OF DRAWINGS

[0013] Figure 1 A flowchart of an individualized ovulation induction scheme based on multi-cycle time sequence modeling and a prediction method for gonadotropin starting dose provided by an embodiment of the application is shown; Figure 2 A structural schematic diagram of a prediction device for an individualized ovulation induction scheme based on multi-cycle time sequence modeling and a prediction method for gonadotropin starting dose provided by an embodiment of the application is shown; Figure 3 A module schematic diagram of a prediction device for an individualized ovulation induction scheme based on multi-cycle time sequence modeling and a prediction method for gonadotropin starting dose provided by an embodiment of the application is shown. DETAILED DESCRIPTION

[0014] The preferred embodiments of the application are described below in conjunction with the accompanying drawings, and it should be understood that the preferred embodiments described herein are only used to illustrate and explain the application, and cannot limit the application.

[0015] The prediction method for an individualized ovulation induction scheme based on multi-cycle time sequence modeling and a prediction method for gonadotropin starting dose according to the exemplary embodiments of the application are described below in conjunction with Figure 1 It should be noted that the following application scenarios are only shown for the purpose of facilitating the understanding of the spirit and principles of the application, and the embodiments of the application are not limited in this respect. On the contrary, the embodiments of the application are applied to any applicable scenario.

[0016] In an embodiment, Figure 1A flowchart of a personalized ovulation induction scheme and prediction method of starting dose of gonadotropin based on multi-cycle timing modeling according to an embodiment of the present application is schematically shown.

[0017] In S101, multi-cycle clinical data of an IVF / ICSI assisted reproduction treatment patient and basic information of a target user are received, and patient samples with reproductive organ malformation, reproductive system tumor or chromosomal abnormality are excluded, to generate an initial patient data set.

[0018] In one embodiment, multi-cycle clinical data of an IVF / ICSI assisted reproduction treatment patient and basic information of a target user are received, and the data covers the full-dimension key information of the patient's complete treatment cycle, ensuring the continuity and integrity of the data. The multi-cycle clinical data includes the relevant data of the patient's 3 COS treatment cycles. In the first cycle (March-April 2022), the antral follicle count (AFC) is 8, the anti-Mullerian hormone (AMH) is 1.2 ng / mL, the follicle-stimulating hormone (FSH) is 7.5 IU / L, the Gn starting dose is 300 IU, and the number of oocytes obtained is 6. In the second cycle (June-July 2022), the AFC is 7, the AMH is 1.1 ng / mL, the FSH is 7.8 IU / L, the Gn starting dose is 325 IU, and the number of oocytes obtained is 5. In the third cycle (September-October 2022), the AFC is 6, the AMH is 1.0 ng / mL, the FSH is 8.0 IU / L, the Gn starting dose is 350 IU, and the number of oocytes obtained is 4. The data also includes the medication regimen of each cycle (all are antagonist regimens), the endometrial morphology (all are type A), the pituitary prolactin (PRL), the anti-Mullerian hormone (AMH), the follicle-stimulating hormone (FSH), the luteotropic hormone (LH), the prolactin (PRL), the progesterone (P), the estradiol (E2), the testosterone (T), and the androstenedione (AND) laboratory test data.

[0019] The target user's basic information is as follows: the patient is 36 years old, the body mass index (BMI) is 34.7 kg / m², the type of infertility is primary infertility, the cause of infertility is unknown, the duration of infertility is 4 years, the number of previous pregnancies is 0, the number of previous deliveries is 0, the anti-Mullerian hormone (AMH) is 1.08 ng / mL, the antral follicle count (AFC) is 8, the basal follicle-stimulating hormone (FSH) is 7.69 IU / L, the luteinizing hormone (LH) is 2.82 IU / L, the pituitary prolactin (PRL) is 13.5 ng / mL, the testosterone (T) is 0.69 ng / dL, the androstenetrione (A) is 2.39 ng / mL, the previous COS scheme is the antagonist scheme, and the number of previous oocytes is 7. A strict sample exclusion process is performed on the received patient data, and patient samples with reproductive organ malformation, reproductive system tumor or chromosomal abnormality are excluded to avoid abnormal samples affecting the accuracy and reliability of model construction. Reproductive organ malformation exclusion: patients with congenital reproductive organ malformation such as uterus septum, double-angle uterus, and absence of fallopian tube are excluded, for example, a patient was diagnosed with complete uterus septum by ultrasound examination, and the related data was included in the exclusion range. Reproductive system tumor exclusion: patients with reproductive system malignant tumors such as ovarian cancer, endometrial cancer, and cervical cancer, as well as patients with ovarian benign tumors larger than 5 cm in diameter that may affect ovarian function, such as a patient with a history of ovarian serous cystadenoma surgery (tumor diameter 6 cm), are excluded. Chromosomal abnormality exclusion: patients diagnosed with chromosomal number or structure abnormalities by chromosomal karyotype analysis are excluded, including any one of the couple with abnormalities, for example, a patient with a chromosomal karyotype of 47, XX, +21 (Down syndrome), whose related treatment data is excluded.

[0020] After data reception and sample exclusion, the multi-cycle clinical data of patients meeting the inclusion criteria and the target user's basic information are structured and integrated to form an initial patient data set. The data set is stored in a unified format specification for subsequent preprocessing and model training. The integrated initial patient data set contains 5000 patient data meeting the standard, each data containing basic information (age, BMI, infertility-related information, etc.), multi-cycle clinical test data (AMH, FSH, LH, etc.), ultrasound examination data (AFC, endometrial morphology, etc.), medication data (COS scheme, Gn dose, etc.), and treatment outcome data (number of oocytes, etc.). All data are stored in the format of "patient unique identifier-cycle number-data type-specific value", such as "P0001-1-AMH-1.2 ng / mL" and "P0001-2-Gn dose-325 IU", to ensure the standardization and traceability of the data.

[0021] S102, the SHAP method is used to screen key prediction factors and quantify feature contribution, multi-cycle clinical data in the initial patient data set is called, the training set and the test set are divided according to the preset processing rule, the pre-stored sample balance optimization algorithm and the model evaluation rule are received, the balanced training sample is generated by processing the training set, the sliding window time sequence processing module and the multi-algorithm model training engine are started, and the patient historical period feature correlation data and the initial individualized ovulation induction strategy recommendation model are generated.

[0022] In an embodiment, the multi-cycle clinical data in the initial patient data set is processed based on the preset feature screening rule, the SHAP method is used to screen key prediction factors and quantify the contribution of each feature, and the core prediction factor set and the feature importance weight information are generated. Based on the preset feature screening rule, the multi-cycle clinical data in the initial patient data set is processed, the prediction factors with a missing proportion of more than 40% are first removed, then the SHAP method (SHapley Additive exPlanations) is used to calculate the contribution of each feature to the prediction result, the key prediction factors are screened and the feature importance weight is quantified, and the core prediction factor set and the feature importance weight information are generated.

[0023] The initial patient data set contains 30 candidate prediction factors including patient age, BMI, AMH (anti-Mullerian hormone), AFC (antral follicle count), FSH (follicle-stimulating hormone), LH (luteinizing hormone), PRL (pituitary prolactin), infertility type, infertility duration, previous COS scheme, and the number of oocytes. After screening, the "androstenedione (AND)" index with a missing proportion of 45% is removed, and the SHAP method is used to calculate that the feature importance weight of AMH is 0.23 (the highest), the AFC is 0.19, the age is 0.17, the FSH is 0.15, the previous number of oocytes is 0.12, and the weight of the remaining indicators is less than 0.08. Finally, AMH, AFC, age, FSH, previous number of oocytes, LH, PRL, infertility type, infertility duration, and previous COS scheme are determined as the core prediction factor set, and the feature importance weight information is stored according to the above values.

[0024] The data set division scheme of the core predictor set is designed based on the model training requirements, the sample size standard of EPV≥20 is followed, the division ratio and screening rules of the training set and the test set are determined, and the data set division scheme is generated; in combination with the sample imbalance condition and the model generalization requirement, the pre-stored SMOTE sample balancing optimization algorithm and the model evaluation rules are called to perform oversampling processing on the training set to generate balanced training samples. The data set division scheme of the core predictor set is designed based on the model training requirements, the sample size standard of EPV≥20 of each independent variable is strictly followed, the division ratio and screening rules of the training set and the test set are determined; in view of the sample imbalance condition, in combination with the model generalization requirement, the pre-stored SMOTE (synthetic minority over-sampling technique) sample balancing optimization algorithm and the model evaluation rules are called to perform oversampling processing on the training set to generate balanced training samples.

[0025] The core predictor set contains 10 independent variables, and at least 200 samples are required for model development according to the standard of EPV≥20, and finally 5469 patient data meeting the requirements are selected from the initial patient data set for COS scheme prediction model construction. The data set division ratio is set as 80% for the training set and 20% for the test set, and the screening rule is to randomly divide according to the time sequence (to avoid data bias), that is, 3632 data from January 1, 2021 to March 31, 2023 are divided into the development set (including 2905 training set and 727 test set), and 1837 data from April 1, 2023 to December 31, 2023 are divided into the external validation set, forming a complete data set division scheme.

[0026] The proportion of "antagonist regimen" samples in the training set is 65%, the proportion of "agonist regimen" is 25%, and the proportion of "micro-stimulation regimen" is only 10%, which is obviously imbalanced. After oversampling the "micro-stimulation regimen" and "agonist regimen" samples by the SMOTE algorithm, the proportion of the three types of regimen samples is adjusted to 55%, 25% and 20% respectively after generating synthetic samples, and balanced training samples are obtained.

[0027] The core predictor set, the data set division scheme, and the balanced training samples are integrated, the sliding window time series processing module and the multi-algorithm model training engine are started, and the patient historical cycle characteristic correlation data and the initial individualized ovulation stimulation strategy recommendation model are generated. The core predictor set, the data set division scheme, and the balanced training samples are integrated, the sliding window time series processing module and the multi-algorithm model training engine (including LightGBM, ExtraTrees, etc.) are started, the time sequence correlation of different cycle data of the patient is captured by the sliding window, the model is constructed by the multi-algorithm engine, and the patient historical cycle characteristic correlation data and the initial individualized ovulation stimulation strategy recommendation model are generated.

[0028] The integrated data includes the complete dimensions of the core predictors, the divided data sets, and the balanced training samples. The sliding window time series processing module sets the window size to "current cycle + previous 1 cycle", extracts the data of the previous cycle of the patient, such as AFC, AMH change, and the effect of the previous COS scheme (number of oocytes obtained), and establishes an association with the current cycle core predictor to generate historical cycle characteristic association data. For example, the association data of patient P0012 is "first cycle AFC 9 - AMH 1.3 ng / mL - antagonist scheme - number of oocytes obtained 7 -> second cycle AFC 8 - AMH 1.2 ng / mL - to be predicted scheme".

[0029] The multi-algorithm model training engine simultaneously calls the LightGBM and ExtraTrees algorithms to set the "recommended COS scheme + Gn starting dose range" as the prediction target, and performs initial model training based on the balanced training samples. The LightGBM algorithm is set to a learning rate of 0.05 and a maximum tree depth of 8, and the ExtraTrees algorithm is set to a decision tree number of 100 and a random feature selection ratio of 0.7. Finally, an initial individualized ovulation induction strategy recommendation model containing a scheme prediction module and a dose preliminary recommendation module is generated.

[0030] In S103, the initial individualized ovulation induction strategy recommendation model is trained using the balanced training samples and the test set, and the training results are output in combination with the model evaluation rules to generate a target ovulation prediction model.

[0031] In one embodiment, the balanced training samples and the test set are split into scheme training data subsets and dose training data subsets according to the prediction target, and are input into the scheme prediction module and the dose preliminary recommendation module of the initial individualized ovulation induction strategy recommendation model, respectively, to perform synchronous iterative training and generate model training process data for each module. The training process of each module incorporates sliding window time series modeling logic. The scheme prediction module optimizes classification parameters using multi-cycle association data including the effect of the patient's previous COS scheme and ovarian response, and the dose prediction module optimizes regression parameters using multi-cycle association data including previous Gn dose and number of oocytes obtained. The dose prediction module splits the data subsets into first to fourth and subsequent cycle subsets and performs independent training of each cycle sub-model. First, the balanced training samples and the test set are split according to the two targets of "COS scheme prediction" and "Gn starting dose prediction", and then the corresponding subsets are input into the two dedicated modules of the model for synchronous iterative training and fusion of time series logic, and the dose module is independently modeled according to the cycle.

[0032] There are 20749 balanced training samples and 5187 test samples. After splitting according to the prediction target, the training data subset of the scheme (including 16599 training samples and 4150 test samples) contains patient age, AMH, and previous COS scheme type classification-related data; the dose training data subset (including 15562 training samples and 3890 test samples) contains previous Gn dose, number of oocytes, AFC, and regression-related data.

[0033] The scheme prediction module training example is as follows: the sliding window time sequence logic is fused (the window is set to “current period + previous 1 period”), and the patient's previous COS scheme effect (such as 5 oocytes in the previous period of antagonist scheme) and multi-period correlation data such as ovarian response (such as follicular development speed 1.2 mm / day) are used to optimize the classification parameters. For example, a certain patient obtained 3 oocytes (average effect) in the first cycle with a stimulator scheme, and when the second cycle data is included in the training, the model adjusts the classification weight and improves the recognition sensitivity of the more suitable scheme.

[0034] The dose prediction module training example is as follows: the sliding window time sequence logic is also fused, and the previous Gn dose (such as 300 IU in the previous period) and multi-period correlation data such as the number of oocytes (such as 6) are used to optimize the regression parameters. At the same time, the data subset is split by period, the first period sub-model uses 18091 training data, the second period uses 4943, the third period uses 650, and the fourth and subsequent periods use 418, which are independently trained. For example, the third period sub-model optimizes the regression coefficient for the characteristics of possible decline in ovarian response, so that the dose prediction is more in line with the physiological characteristics of the period.

[0035] Combined with the preset model evaluation rules, the training process data of the double module is quantitatively evaluated, and a model performance evaluation report containing the double module and each cycle sub-model is generated; wherein, the evaluation covers internal five-fold cross-validation and external independent data set validation, the scheme prediction module takes accuracy as the core evaluation index, and the dose prediction module takes R² value as the core evaluation index. The internal five-fold cross-validation and external independent data set validation are combined to quantitatively evaluate the training process data of the double module according to the preset rules, and the performance evaluation report is generated, and the core indicators have their own emphasis. The internal five-fold cross-validation splits the scheme training set 16599 into 5 groups according to the ratio of 1:4, each group has 3320 as the validation set and 13279 as the training set, and the cycle is verified; the external independent data set is 1837 patient data from April to December 2023, which is used for model generalization test alone.

[0036] The scheme prediction module evaluation example is as follows: the core index is accuracy, after a certain round of training, the internal five-fold cross-validation average accuracy is 75%, the external validation set accuracy is 73%, the report records “the first cycle scheme prediction accuracy is 76%, the second cycle is 74%, the third and subsequent cycles are 71%”, and the classification error distribution of each cycle is labeled.

[0037] The dose prediction module evaluates as follows, and the core indicator is R2 value. The R2 value of the first cycle sub-model training set is 0.79, the R2 value of the test set is 0.79, and the R2 value of the external validation set is 0.76. The R2 value of the third cycle sub-model training set is 0.62, the R2 value of the test set is 0.65, and the R2 value of the external validation set is 0.59. The goodness of fit and error sources of each cycle sub-model are explicitly reported in the report.

[0038] In combination with the preset model evaluation rules, the training process data of the double modules are quantitatively evaluated to generate a model performance evaluation report containing the double modules and each cycle sub-model. The evaluation covers internal five-fold cross-validation and external independent data set validation. The accuracy rate is the core evaluation indicator for the scheme prediction module, and the R2 value is the core evaluation indicator for the dose prediction module. For the scheme prediction module, the LightGBM algorithm is selected for the first cycle (training set 18091 cases, large data volume) and the second cycle (training set 4943 cases). The ExtraTrees algorithm is selected for the third cycle (training set 650 cases), the fourth cycle and subsequent cycles (training set 418 cases). For the dose preliminary recommendation module, the LightGBM algorithm is selected for the first cycle (R2 target ≥ 0.75) and the second cycle (R2 target ≥ 0.63). The ExtraTrees algorithm is selected for the third cycle (R2 target ≥ 0.58) and the fourth cycle and subsequent cycles (R2 target ≥ 0.55).

[0039] The initial parameters of the LightGBM algorithm (taking the first cycle dose sub-model as an example) are learning rate 0.05, maximum tree depth 8, and leaf node number 20. After training, the R2 value is 0.77, and the loss value is 0.25. Through iteration adjustment by the Optuna tool, when the learning rate is reduced to 0.03, the maximum tree depth is increased to 10, and the leaf node number is adjusted to 30, the R2 value is increased to 0.79, and the loss value is reduced to 0.2, reaching the optimal state, and there is no overfitting phenomenon (the external verification R2 value is 0.76, and the difference with the internal verification is ≤3%).

[0040] The initial parameters of the ExtraTrees algorithm (taking the third cycle scheme sub-model as an example) are decision tree number 100, random feature selection proportion 0.6, and minimum sample split number 5. After training, the accuracy rate is 70%, and the loss value is 0.32. After adjustment to decision tree number 200, random feature selection proportion 0.7, and minimum sample split number 3, the accuracy rate is increased to 72%, the loss value is reduced to 0.28, the external verification accuracy rate is 69%, and the clinical standard is met.

[0041] The optimization results are as follows: for the regimen prediction module, the accuracy of each cycle is ≥70% (external validation of the fourth and subsequent cycles is 68%, close to the standard), and the classification error is reduced by 15-20% compared to before optimization. For the dose preliminary recommendation module, the R² value of each cycle is ≥0.57 (external validation), and the residual standard deviation is reduced by 8-12% compared to before optimization, achieving optimal performance.

[0042] Integrate the optimized parameters of the dual module, the parameter configuration of each cycle sub-model, and the performance evaluation standard results, clarify the synergistic linkage logic of the COS regimen prediction module and the multi-cycle gonadotropin dose prediction module, build a complete model architecture, and generate the target ovulation prediction model. From the module dimension and cycle dimension, respectively, comb and integrate the optimized parameters to form a unified parameter configuration system, ensuring that the parameters of each module and the corresponding cycle sub-model are traceable and callable.

[0043] The parameters of the regimen prediction module are integrated as follows: for the first cycle, the LightGBM algorithm is used with the hyperparameter configuration of learning rate 0.05, maximum tree depth 8, and leaf node number 25. Under this configuration, the external validation accuracy is 75%, which can accurately identify patients suitable for antagonist regimen and agonist regimen. For the second cycle, the LightGBM algorithm is used with the hyperparameter adjustment of learning rate 0.04, maximum tree depth 9, and leaf node number 20. The external validation accuracy is 73%, which can optimize the scheme selection logic for patients with slight changes in ovarian response in the second cycle. For the third cycle, the ExtraTrees algorithm is used with the hyperparameters of decision tree number 200, random feature selection proportion 0.7, and minimum sample split number 3. The external validation accuracy is 69%, which is suitable for patients with less sample size and possibly decreased ovarian reserve in the third cycle, reducing the risk of overfitting. For the fourth and subsequent cycles, the ExtraTrees algorithm is used with the hyperparameters of decision tree number 180, random feature selection proportion 0.65, and minimum sample split number 4. The external validation accuracy is 68%, which can adapt to long-term treatment patients with decreased ovarian response stability.

[0044] The dose preliminary recommendation module parameters are integrated as follows, the first cycle: using LightGBM algorithm, the hyperparameters are learning rate 0.03, maximum tree depth 10, leaf node number 30, external validation R² value 0.76, which can accurately fit the nonlinear relationship between AMH, AFC and Gn dose, such as patients with AMH 1.2 ng / mL and AFC 9, the prediction dose error can be controlled within 20 IU. The second cycle: using LightGBM algorithm, the hyperparameters are learning rate 0.04, maximum tree depth 9, leaf node number 25, external validation R² value 0.63, for the subtle changes of ovarian response of patients in the second cycle, the dose regression coefficient is adjusted, such as patients with previous oocyte number 5, the dose prediction is more in line with their historical treatment response. The third cycle: using ExtraTrees algorithm, the hyperparameters are decision tree number 220, random feature selection ratio 0.7, minimum sample number 3, external validation R² value 0.59, combined with the clinical characteristics of the decline of ovarian reserve in the third cycle, the dose adjustment range is optimized, such as patients with AFC from 8 to 6, the dose recommendation can be appropriately increased by 25-50 IU. The fourth and subsequent cycles: using ExtraTrees algorithm, the hyperparameters are decision tree number 200, random feature selection ratio 0.65, minimum sample number 4, external validation R² value 0.57, for the individual differences of long-term treatment patients, the dynamic adjustment range of dose recommendation is expanded, and the prediction error is reduced through multiple decision tree integration.

[0045] According to the principle of "stratified design, function closed loop", a model architecture containing 5 core levels is built to ensure that each level has clear responsibilities and efficient data flow. The data input layer receives the basic information of the target user (age, BMI, infertility type, infertility duration, etc.), clinical test information (AMH, FSH, LH, PRL, etc.), ultrasound examination information (AFC, endometrial morphology), and previous treatment information (previous COS scheme, previous oocyte number), supports batch import and single input of data, and automatically detects data integrity (such as missing key indicators such as AMH and AFC, prompting "core ovarian reserve indicators need to be supplemented").

[0046] The feature processing layer cleanses the input data (corrects input errors, such as correcting BMI 347 kg / m² to 34.7 kg / m²), standardizes the data (unifies the units of indicators, such as converting testosterone 0.69 ng / dL and androstenedione 2.39 ng / mL according to the preset format), calculates the contribution of each feature through SHAP method, selects core predictors (such as AMH, AFC, age, etc.), eliminates redundant features with contribution less than 0.05 (such as previous number of deliveries, which has a contribution of only 0.03 in dose prediction), and generates a feature set that can be directly used for prediction.

[0047] The double prediction module layer includes a scheme prediction module and a dose preliminary recommendation module, and the two modules are linked in real time through a data interface. The scheme prediction module calls the corresponding sub-model according to the cycle information, and outputs the COS scheme and the matching degree score; the dose module calls the exclusive sub-model based on the scheme result and the cycle information to output the Gn starting dose and the adjustment suggestion, such as “recommended antagonist scheme (matching degree 85 points), Gn starting dose 356.66 IU (recommended range 350-365 IU)”.

[0048] The evaluation calibration layer calls the clinical applicability verification data (such as the past treatment effect of patients similar to the target user characteristics-AMH1.0-1.2 ng / mL, AFC7-9 patients using an antagonist scheme, and the average number of oocytes 6-8), and compares the fitting degree of the current prediction result and the clinical data; if there is a deviation between the expected number of oocytes (6) corresponding to the predicted dose and the average number of oocytes (7) of the similar clinical cases, a dose calibration coefficient is automatically generated, and the recommended range is adjusted to 355-370 IU, to ensure that the prediction result is more consistent with the real clinical effect.

[0049] The result output layer outputs the final recommended result in the form of a “clinical report”, which includes three core contents: first, the individualized COS scheme, which clearly recommends the scheme type (such as an antagonist scheme) and the recommended basis (such as “combined with your AMH1.08 ng / mL, and the past antagonist scheme oocyte number 7, the scheme matching degree is 85 points”); second, the Gn starting dose, which marks the specific dose value (356.66 IU), the recommended range (350-365 IU), and the adjustment basis (such as “AFC8, which is 1 less than the first cycle, and the dose is increased by 25 IU compared with the previous cycle”); and third, the clinical monitoring suggestion, such as “monitor the LH level on the 3rd day of medication, the follicle size and estradiol level on the 5th day, and if the follicle development is slow, increase the dose by 25 IU per day”, to help doctors achieve dynamic diagnosis and treatment management.

[0050] The model takes a web calculator as a carrier, and doctors access the tool through a browser, manually input the patient's cycle number, AFC, age and other key information (or upload standardized electronic medical record data), click the “predict” button, and the system completes the calculation within 3-5 seconds, outputs a visual recommendation report, supports report download and printing, and is directly used for clinical diagnosis and treatment records.

[0051] S104, based on the hierarchical classification strategy and the weighted integration rule, the balanced training sample and the test set evaluation result are processed to generate a COS scheme classification prediction signal and a multi-cycle Gn dose prediction adjustment information.

[0052] In an embodiment, the balanced training samples are hierarchically processed based on a hierarchical classification strategy to generate a COS scheme classification candidate signal, wherein the hierarchical classification first trains a binary classification model for the COS scheme category with the largest sample size and adjusts the decision probability threshold, then trains a multi-classification model for the remaining categories, and combines SMOTE oversampling to solve the sample imbalance. Based on the hierarchical classification strategy, the balanced training samples are hierarchically processed, and a binary classification model is trained for the COS scheme category with the largest sample size (such as the antagonist scheme), and the decision probability threshold is adjusted (such as setting the threshold to 0.65) to improve the recognition accuracy of this category; then a multi-classification model is trained for the remaining categories (agonist scheme, micro-stimulation scheme), and SMOTE oversampling technology is combined before training to further balance the sample distribution of the remaining categories, and finally a COS scheme classification candidate signal containing the prediction probability of each COS scheme is generated.

[0053] In the balanced training samples, the antagonist scheme is first taken as the positive class and the other schemes are taken as the negative class to train a binary classification model, and after adjusting the decision probability threshold to 0.65, the recognition accuracy of the antagonist scheme by the model reaches 88%; then the agonist scheme and micro-stimulation scheme samples are SMOTE oversampled (the proportion of the two types of samples after expansion is 50%), and a multi-classification model is trained, which has a distinction accuracy of 82% for the two types of schemes. The final generated COS scheme classification candidate signal is presented in the form of "scheme type-prediction probability", for example, the candidate signal for patient P0035 is "antagonist scheme-0.89, agonist long scheme-0.71, agonist ultra-long scheme-0.61, agonist short scheme-0.52, agonist ultra-short scheme-0.48, micro-stimulation scheme-0.60".

[0054] The test set evaluation results and the classification candidate signal are fused and optimized using a weighted integration rule to generate a COS scheme classification prediction signal, wherein the weighted integration assigns different weights to the prediction results of the binary classifiers corresponding to different categories to improve the prediction accuracy of minority categories. The test set evaluation results and the COS scheme classification candidate signal are fused and optimized using a weighted integration rule, different weights are assigned to different categories of samples according to the sample proportion and model prediction accuracy (the weight of the category with less sample size and lower accuracy is higher), the prediction results of the two types of data are combined by weighted voting to generate the final COS scheme classification prediction signal, and the prediction accuracy of minority categories is improved.

[0055] The COS schemes are classified as follows: antagonist scheme (majority class, most widely used in clinical application), agonist long scheme (agonist subtype 1, suitable for patients with good ovarian reserve), agonist super long scheme (agonist subtype 2, suitable for patients with endometriosis-related infertility), agonist short scheme (agonist subtype 3, suitable for patients with sensitive ovarian response), agonist super short scheme (agonist subtype 4, suitable for patients with slightly low ovarian reserve), and micro-stimulation scheme (minority class, suitable for elderly patients and patients with low ovarian response).

[0056] Based on the principle that "the less the sample size, the higher the priority of clinical adaptation, and the higher the weight", the weight distribution rules are re-set to ensure that the minority class scheme can be accurately identified. Specifically, the micro-stimulation scheme (minority class): weight 0.25; the agonist super long scheme (with less sample size in the subtype): weight 0.20; the agonist super short scheme (with less sample size in the subtype): weight 0.18; the agonist long scheme (with medium sample size in the subtype): weight 0.15; the agonist short scheme (with medium sample size in the subtype): weight 0.12; and the antagonist scheme (majority class, with the largest sample size): weight 0.10.

[0057] An example of the adjusted weighted calculation is as follows: the test set evaluation result of a certain patient is as follows (the numerical value represents the scheme adaptation probability): the test set evaluation result is micro-stimulation scheme-0.61; the classification candidate signal is antagonist scheme-0.58, agonist long scheme-0.52, agonist super long scheme-0.49, agonist short scheme-0.45, agonist super short scheme-0.43, and micro-stimulation scheme-0.32.

[0058] According to the adjusted weight rules, the weighted scores of each scheme are calculated respectively: for the micro-stimulation scheme, the test set evaluation result score (0.61x0.25) + the classification candidate signal score (0.32x0.25) = 0.1525 + 0.08 = 0.2325. For the agonist super long scheme, the classification candidate signal score (0.49x0.20) = 0.098 (when there is no separate test set evaluation result, only the candidate signal score is calculated). For the agonist super short scheme, the classification candidate signal score (0.43x0.18) = 0.0774. For the agonist long scheme: the classification candidate signal score (0.52x0.15) = 0.078. For the agonist short scheme: the classification candidate signal score (0.45x0.12) = 0.054. For the antagonist scheme: the classification candidate signal score (0.58x0.10) = 0.058.

[0059] According to the weighted score ranking, the micro-stimulation scheme has the highest score of 0.2325 and becomes the optimal adaptation scheme, and the final prediction signal is generated as "micro-stimulation scheme-0.2325", successfully identifying the adaptation needs of the patient for the minority class scheme, which is consistent with the original logic and more accurate in classification.

[0060] The evaluation results of the equalized training samples and the test set are split according to the COS cycle stratification rule to generate cycle-specific training data. The equalized training samples and the test set evaluation results are split according to the COS cycle stratification rule (first cycle, second cycle, third cycle, fourth and later cycle) to ensure that each cycle of samples contains complete core predictor data, program execution records and treatment outcomes (such as the number of oocytes), and generates cycle-specific training data to provide targeted data support for multi-cycle Gn dose prediction sub-model construction.

[0061] From the equalized training samples (20749 cases), 18091 cases of first cycle training data, 4943 cases of second cycle, 650 cases of third cycle, and 418 cases of fourth and later cycle are split; from the test set (5187 cases), 6389 cases, 1757 cases, 1831 cases, and 1071 cases of corresponding cycle test data are split. Each cycle-specific training data contains core predictors such as AMH, AFC, age, FSH, and corresponding Gn dose, number of oocytes, etc. For example, in the third cycle-specific training data, patient P0042's record is "AMH 0.8 ng / mL, AFC 5, age 35, FSH 9.2 IU / L, Gn dose 375 IU, number of oocytes 3".

[0062] Based on the cycle-specific training data, a multi-cycle Gn dose prediction sub-model is constructed, and the model parameters are optimized according to the differences in clinical characteristics of each cycle to generate multi-cycle Gn dose prediction adjustment information. Based on the cycle-specific training data, a multi-cycle Gn dose prediction sub-model is constructed, and the model parameters are optimized according to the differences in clinical characteristics of each cycle (such as the first / second cycle with relatively good ovarian reserve, and the third and later cycle with possibly declining ovarian response). By learning the correlation between core predictors and Gn dose in each cycle, multi-cycle Gn dose prediction adjustment information containing Gn dose prediction values and adjustment amplitudes in each cycle is generated.

[0063] A LightGBM sub-model is constructed for the first cycle, and the optimized parameters are learning rate 0.03 and maximum tree depth 10. The prediction R² value of this model for the first cycle Gn dose is 0.79. An ExtraTrees sub-model is constructed for the third cycle, and the parameters are decision tree number 200 and random feature selection ratio 0.7. The prediction R² value is 0.65. An example of the generated multi-cycle Gn dose prediction adjustment information is as follows: patient P0042 is in the third cycle of treatment, the core predictors are "AMH 0.8 ng / mL, AFC 5, age 35", the model predicts the Gn dose to be 380 IU, and the adjustment information is "third cycle - Gn starting dose 380 IU - recommended adjustment amplitude ± 25 IU".

[0064] S105, process the COS scheme classification prediction signal, multi-cycle Gn dose prediction adjustment information and clinical applicability verification data to generate individualized scheme matching degree score and dose recommendation accuracy calibration information.

[0065] In an embodiment, the COS scheme classification prediction signal, multi-cycle Gn dose prediction adjustment information and clinical applicability verification data are retrieved, and a three-dimensional evaluation dimension is established, wherein the clinical applicability verification data includes real clinical data such as the execution effect of the previous COS scheme, ovary response feedback and complication occurrence. The COS scheme classification prediction signal, multi-cycle Gn dose prediction adjustment information are retrieved, and the clinical applicability verification data is collected at the same time, and a "prediction signal-dose adjustment-clinical verification" three-dimensional evaluation dimension is established. The clinical applicability verification data is real clinical diagnosis and treatment data, including the execution effect of the previous COS scheme (such as the number of oocytes, embryo quality), ovary response feedback (such as follicle development speed, hormone level change) and complication occurrence (such as the incidence of ovarian hyperstimulation syndrome (OHSS)).

[0066] The COS scheme classification prediction signal is "antagonist scheme-0.89, agonist long scheme-0.71, agonist super long scheme-0.61, agonist short scheme-0.52, agonist super short scheme-0.48, and micro-stimulation scheme-0.60", and the multi-cycle Gn dose prediction adjustment information is "third cycle-Gn starting dose 380 IU-suggested adjustment range ± 25 IU".

[0067] The clinical applicability verification data retrieves the patient's previous third cycle treatment record: micro-stimulation scheme is adopted, Gn starting dose is 350 IU, the number of oocytes is 3, OHSS does not occur, and the follicle development speed meets the expectation; at the same time, the treatment data of patients with similar clinical characteristics are matched, such as patients with AMH 0.8 ng / mL and AFC 5, the number of oocytes of patients using the antagonist scheme is an average of 4, and the OHSS incidence is 1.2%.

[0068] Based on the preset scheme adaptation threshold and dose-response matching benchmark, the prediction signal, dose adjustment information and clinical verification data are compared to identify the correction signal of scheme adaptation deviation and dose-response mismatch, wherein the scheme adaptation deviation is determined by whether the fitting degree of the predicted scheme and the clinical optimal scheme meets the standard, and the dose-response mismatch is determined by whether the number of oocytes corresponding to the predicted dose and the ovary response meet the expected benchmark. Based on the preset scheme adaptation threshold and dose-response matching benchmark, the prediction signal, dose adjustment information and clinical verification data are compared. The scheme adaptation deviation is determined by whether the fitting degree of the predicted scheme and the clinical optimal scheme meets the standard (the adaptation threshold is set to 0.6), and the dose-response mismatch is determined by whether the number of oocytes corresponding to the predicted dose and the ovary response meet the expected benchmark, and finally the deviation is identified and the correction signal is generated.

[0069] The scheme adaptation threshold is set to 0.6, the antagonist scheme fit degree in the prediction signal is 0.89 (up to standard), but the clinical verification data shows that the oocyte number of patients with similar characteristics using the antagonist scheme is better, and the patient's previous oocyte number using the micro-stimulation scheme is less, it is determined that there is a "prediction scheme and individual optimal scheme adaptation deviation", and the correction signal is "preferentially recommend the antagonist scheme, and need to match the adaptation dose".

[0070] The dose-response matching benchmark is "Gn dose 380 IU corresponding to oocyte number 4-6", combined with the patient's previous 350 IU oocyte number of 3, the prediction dose of 380 IU meets the dose-response positive correlation rule, it is determined that there is no dose-response mismatch, and no dose correction signal needs to be generated.

[0071] Collect scheme matching data and dose calibration cases under different clinical scenarios, construct a scheme-dose-clinical effect correlation model, calculate the adaptation coefficient of the current prediction parameter node using the model, and generate individualized scheme matching degree score and dose recommendation precision calibration information using the coefficient. Collect scheme matching data and dose calibration cases under different clinical scenarios (such as patients' treatment data at different AMH levels and age stratification), construct a scheme-dose-clinical effect correlation model, calculate the adaptation coefficient of the current prediction parameter node using the model, and generate individualized scheme matching degree score and dose recommendation precision calibration information using the coefficient.

[0072] Collect 1000 cases of different clinical scenario data, covering patients with AMH 0.5-2.0 ng / mL, age 25-40 years old, different COS cycles, after constructing the correlation model, input the patient's "third cycle, AMH 0.8 ng / mL, AFC 5, predicted antagonist scheme, dose 380 IU" and other parameters, and calculate the adaptation coefficient 0.85.

[0073] Based on the adaptation coefficient, the individualized scheme matching degree score is 85 points (full score 100 points), and the dose recommendation precision calibration information is "Gn starting dose 380 IU-calibration recommended range 370-390 IU-recommendation precision level A", which clearly shows the adaptation and precision of the scheme and dose to the patient.

[0074] S106, clean and standardize the target user's basic information to generate user input data that can be used for prediction.

[0075] In one implementation, the initial basic information of the target user P0064 is "age 36, BMI 34.7 kg / m², secondary infertility type, no infertility cause filled in, infertility duration 4 years, previous pregnancy 1 time, and previous delivery 0 times". Among them, "infertility cause" is the core basic information, which is supplemented as "tubal factor" by checking the user's electronic medical record. If the key information such as "age" and "BMI" is missing, it is directly determined that the user's basic information is incomplete and needs to be supplemented before entering the next step of processing.

[0076] Exclusion item compliance verification: Verify whether the target user P0064 has exclusion conditions: ① Check the ultrasound examination report, there is no record of reproductive organ malformation such as uterine septum and bicornuate uterus; ② Retrieve medical history, there is no history of reproductive system tumors such as ovarian cancer and endometrial cancer, and no record of benign ovarian tumors (diameter > 5 cm); ③ Chromosome karyotype analysis result is 46, XX (normal), both husband and wife have no chromosomal abnormalities, it is determined that the user meets the inclusion criteria, and the basic information is retained; if a user has "ovarian serous cystadenoma surgery history (tumor diameter 6 cm)", the basic information is directly excluded and does not participate in subsequent processing.

[0077] Data accuracy correction: In the initial basic information of the target user P0065, "BMI is recorded as 34.7 kg / m²", which is verified as an input error (actual height 1.65 m, weight 65 kg, calculated BMI 23.9 kg / m²), corrected to the correct value; "infertility duration is recorded as 6 years", which is inconsistent with the time difference (4 years) between "2019 diagnosed with infertility and 2023 treatment" in the electronic medical record, corrected to "4 years", ensuring that the basic information is consistent with the actual diagnosis and treatment record.

[0078] For the target user basic information after cleaning, standardized processing is performed according to "data format unification, classification information coding, and exclusion item identification" to ensure that the information format is unified and can be directly parsed by the model, and completely corresponds to the dimension of the initial basic information. Data format unification: ① Age is unified to retain integers, such as "36 years old" of target user P0064 remains unchanged, if it is "32.5 years old", it is corrected to "33 years old"; ② BMI is unified to retain 1 decimal place, such as "34.7 kg / m²" remains unchanged, "24 kg / m²" is supplemented to "24.0 kg / m²"; ③ Infertility duration, previous pregnancy times, and previous delivery times are unified to retain integers, such as "4 years", "1 time", and "0 times" are all in integer format.

[0079] Coding of classification information: ① Infertility type: "primary infertility" is coded as "0", "secondary infertility" is coded as "1", and the "secondary infertility" of the target user P0064 is coded as "1"; ② Infertility cause: "tubal factor" is coded as "0", "ovulation disorder" is coded as "1", "ovarian factor" is coded as "2", and "other factors" is coded as "3". The "tubal factor" of the target user P0064 is coded as "0". The coding rule is fixed to ensure that the classification basic information of all target users is converted according to the unified standard.

[0080] For the target users who meet the inclusion criteria after verification, "exclusion item identifier-0" (0 represents no exclusion item, meets the inclusion) is added in the basic information. If a user is excluded due to "chromosome karyotype 47, XX, +21", the identifier is "exclusion item identifier-1" (1 represents the existence of exclusion items, does not meet the inclusion), which facilitates the subsequent rapid screening of qualified user basic information. After integrating and cleaning the target user basic information, the structured user input data is generated according to the corresponding relationship of "initial basic information dimension-standardization result", ensuring that each data item can be traced back to the initial basic information, with no new or missing dimensions, which can be directly imported into the target ovulation prediction model.

[0081] As shown in Figure 2 , the final predictable user input data of the target user P0064 is: 'age: 36 (integer, unit: years), BMI: 34.7 (1 decimal place, unit: kg / m²), AMH: 1.08 (2 decimal places, unit: ng / mL), AFC: 8 (integer, unit: pieces), FSH: 7.69 (2 decimal places, unit: IU / L), LH: 2.82 (2 decimal places, unit: IU / L), PRL: 13.5 (1 decimal place, unit: ng / mL), T: 0.69 (2 decimal places, unit: ng / dL), A: 2.39 (2 decimal places, unit: ng / mL), infertility type: 0 (primary infertility), infertility cause: 3 (unknown reason), infertility duration: 4 (integer, unit: years), previous pregnancy times: 0 (integer, unit: times), previous delivery times: 0 (integer, unit: times), previous COS scheme: 0 (antagonist scheme), previous oocyte number: 7 (integer, unit: pieces), exclusion item identifier: 0 (no exclusion situation)'.

[0082] S107, based on the target ovulation prediction model, the user input data that can be used for prediction and the individualized scheme matching degree score and dose recommendation accuracy calibration information are processed to generate and output the individualized COS scheme and Gn starting dose recommendation result.

[0083] In an embodiment, the core prediction logic of the target ovulation prediction model is extracted, and the user input data available for prediction, the individualized scheme matching score, and the dose recommendation accuracy calibration information are collected, and the feature-score-calibration association mapping relationship of the three is established, and the user feature-adaptation score-dose calibration association data is generated. First, the core prediction logic of the target ovulation prediction model is extracted, and the core logic framework of the model for realizing COS scheme and Gn dose prediction based on core prediction factors (age, BMI, AMH, AFC, etc.) is determined; then, the user input data available for prediction (basic information after cleaning and standardization), the individualized scheme matching score (adaptation score based on clinical verification), and the dose recommendation accuracy calibration information (calibrated dose range and accuracy level) are collected, and the association mapping relationship of the three is established, and the user feature-adaptation score-dose calibration association data is generated.

[0084] The core prediction logic of the target ovulation prediction model is to predict the optimal COS scheme and Gn dose based on AMH, AFC, age, etc. through the LightGBM algorithm for the first / second cycle user; and to predict through the ExtraTrees algorithm for the third / fourth and subsequent cycles. The collected user input data (user P0066, first cycle): age 36 years old, BMI 34.7 kg / m², AMH 1.08 ng / mL, AFC 8, FSH 7.69 IU / L, LH 2.82 IU / L, PRL 13.5 ng / mL, T 0.69 ng / dL, A 2.39 ng / mL, infertility type 0 (primary infertility), infertility cause 3 (unknown reason), number of previous pregnancies 0 times, number of previous deliveries 0 times, previous COS scheme 0 (antagonist scheme), number of previous oocytes 7.

[0085] Based on the user feature-adaptation score-dose calibration association data, for the clinical individual characteristics of the target user, the deviation value of the matching score from the optimal adaptation threshold is calculated in real time, combined with the dose calibration information, to generate individualized adjustment parameters, wherein the individualized adjustment parameters include the COS scheme type optimization direction, the Gn starting dose adjustment amplitude, and the applicable cycle, and the parameter generation is synchronized with the user multi-cycle clinical data update frequency. Based on the user feature-adaptation score-dose calibration association data, focusing on the clinical individual characteristics (such as the number of cycles, ovarian reserve level, and previous treatment response) of the target user, the deviation value of the individualized scheme matching score from the optimal adaptation threshold (preset as 80 points) is calculated in real time, combined with the dose recommendation accuracy calibration information, to generate individualized adjustment parameters including the COS scheme type optimization direction, the Gn starting dose adjustment amplitude, and the applicable cycle, and the parameter generation frequency is synchronized with the user multi-cycle clinical data update frequency (such as updating once for every additional treatment cycle).

[0086] Target user P0066 (second cycle) clinical individual characteristics: moderate ovarian reserve (AMH 1.2 ng / mL, AFC 8), no adverse reactions to previous COS treatment, and current treatment is the second cycle. Calculate the bias value: regimen matching score 85 - optimal fit threshold 80 = 5 (positive bias), determine that the current predicted regimen is well adapted and does not need to be adjusted.

[0087] Generate adjustment parameters combined with dose calibration information: COS regimen type optimization direction "maintain antagonist regimen", Gn starting dose adjustment range "reduce 5 IU based on calibration range", applicable cycle "second cycle", and the final adjustment parameters are "antagonist regimen - 315 IU - second cycle". If the user completes the second cycle of treatment later, new third cycle clinical data (such as AMH 1.1 ng / mL, AFC 7) are added, and the adjustment parameters are updated simultaneously to adapt the model algorithm to the clinical characteristics of the third cycle.

[0088] According to the user's COS treatment cycle stratification, combined with the corresponding ovarian response characteristics and model algorithm adaptation differences of each cycle, the cycle-specific recommendation strategy is generated by adjusting the recommendation results of each cycle. The cycle-specific recommendation strategy specifies the optimal COS regimen, Gn starting dose range, and clinical monitoring focus for each cycle. According to the user's COS treatment cycle stratification (first to fourth and subsequent cycles), combined with the corresponding ovarian response characteristics (first / second cycle ovarian reserve is relatively stable, and third and subsequent cycles may have decreased response) and model algorithm adaptation differences (first / second cycle LightGBM, third / fourth and subsequent cycles ExtraTrees), the prediction recommendation results of each cycle are adjusted to adapt to the cycle, and the optimal COS regimen, Gn starting dose range, and clinical monitoring focus for each cycle are specified to generate a cycle-specific recommendation strategy.

[0089] First cycle-specific recommendation strategy (user P0067, first cycle, AMH 1.5 ng / mL, AFC 10): the optimal COS regimen is an antagonist regimen, the Gn starting dose range is 300-320 IU, and the clinical monitoring focus is "monitoring follicular development speed and E2 level on day 5 of medication".

[0090] Third cycle-specific recommendation strategy (user P0068, third cycle, AMH 0.9 ng / mL, AFC 6): the optimal COS regimen is a micro-stimulation regimen, the Gn starting dose range is 275-295 IU, and the clinical monitoring focus is "monitoring follicle number and hormone level daily from day 3 of medication, and adjusting the dose in a timely manner".

[0091] The optimal COS scheme for the fourth and subsequent cycle-specific recommendation strategy (user P0069, fourth cycle, AMH 0.8 ng / mL, AFC 5) is a micro-stimulation scheme, the Gn starting dose range is 260-280 IU, and the clinical monitoring focus is "combined with the previous cycle ovarian response, increase the ultrasound monitoring frequency, and be alert to the low ovarian response".

[0092] In one embodiment, as shown in FIG. 1, Figure 3 The application also provides a device for predicting an individualized ovulation induction scheme and a gonadotropin starting dose based on multi-cycle time series modeling, comprising: The acquisition module 301 is configured to receive multi-cycle clinical data of an IVF / ICSI assisted reproduction treatment patient and target user basic information, exclude patient samples with reproductive organ malformation, reproductive system tumor or chromosomal abnormality, and generate an initial patient data set; The processing module 302 is configured to filter key prediction factors and quantify feature contributions by using the SHAP method, call the multi-cycle clinical data in the initial patient data set, divide the training set and the test set according to a preset processing rule, receive a pre-stored sample balancing optimization algorithm and a model evaluation rule, process the training set to generate a balanced training sample, start a sliding window time series processing module and a multi-algorithm model training engine, generate patient historical cycle feature association data and an initial individualized ovulation induction strategy recommendation model; train the initial individualized ovulation induction strategy recommendation model by using the balanced training sample and the test set, output the training result in combination with the model evaluation rule, generate a target ovulation prediction model; process the balanced training sample, the test set evaluation result based on a hierarchical classification strategy and a weighted integration rule, generate a COS scheme classification prediction signal and multi-cycle Gn dose prediction adjustment information; process the COS scheme classification prediction signal, the multi-cycle Gn dose prediction adjustment information and the clinical applicability verification data, generate an individualized scheme matching degree score and dose recommendation precision calibration information; clean and standardize the target user basic information to generate user input data that can be used for prediction; process the user input data that can be used for prediction, the individualized scheme matching degree score and the dose recommendation precision calibration information based on the target ovulation prediction model, generate and output an individualized COS scheme and a Gn starting dose recommendation result.

[0093] The computer-readable storage medium provided by the above embodiments of the application and the prediction method of the individualized ovulation induction scheme and the gonadotropin starting dose based on multi-cycle time series modeling provided by the embodiments of the application have the same beneficial effects as the methods adopted, run or implemented by the application programs stored therein.

[0094] The various embodiments in the present application are described in a related manner, and the same or similar parts among the various embodiments can be referred to each other. Each embodiment focuses on the difference from other embodiments. In particular, for the prediction method, electronic device, electronic equipment, and readable storage medium embodiments for evaluating individualized controlled ovarian hyperstimulation protocol and gonadotropin starting dose based on multi-cycle time series modeling, since they are basically similar to the above-mentioned embodiments of the prediction method based on multi-cycle time series modeling for individualized controlled ovarian hyperstimulation protocol and gonadotropin starting dose, the description is relatively simple, and the relevant parts can be referred to the above-mentioned part of the description of the embodiments of the prediction method based on multi-cycle time series modeling for individualized controlled ovarian hyperstimulation protocol and gonadotropin starting dose.

Claims

1. A method for predicting individualized ovulation induction protocols and initial gonadotropin doses based on multi-cycle time-series modeling, characterized in that, include: Receive multi-cycle clinical data and target user basic information from IVF / ICSI assisted reproductive treatment patients, exclude patient samples with reproductive organ malformations, reproductive system tumors or chromosomal abnormalities, and generate an initial patient data set; The SHAP method is used to screen key predictive factors and quantify feature contribution. Multi-cycle clinical data from the initial patient dataset are called up. The training set and test set are divided according to the preset processing rules. The pre-stored sample balancing optimization algorithm and model evaluation rules are received. The training set is processed to generate balanced training samples. The sliding window time series processing module and multi-algorithm model training engine are started to generate patient historical cycle feature association data and initial individualized ovulation induction strategy recommendation model. The initial individualized ovulation induction strategy recommendation model was trained using balanced training samples and test sets. The training results were then combined with model evaluation rules to generate a target ovulation induction prediction model. Based on hierarchical classification strategy and weighted integration rules, the evaluation results of balanced training samples and test sets are processed to generate COS scheme classification prediction signals and multi-cycle Gn dose prediction adjustment information. The COS protocol classification prediction signal, multi-cycle Gn dose prediction adjustment information and clinical applicability verification data are processed to generate individualized protocol matching score and dose recommendation accuracy calibration information; The basic information of the target users is cleaned and standardized to generate user input data that can be used for prediction. Based on the target ovulation induction prediction model, the user input data that can be used for prediction and the individualized protocol matching score and dose recommendation accuracy calibration information are processed to generate and output the individualized COS protocol and Gn starting dose recommendation results.

2. The method as described in claim 1, characterized in that, The SHAP method is used to screen key predictive factors and quantify feature contributions. Multi-cycle clinical data from the initial patient dataset is retrieved, and training and test sets are divided according to preset processing rules. Pre-stored sample balancing optimization algorithms and model evaluation rules are received. The training set is processed to generate balanced training samples. The sliding window time series processing module and multi-algorithm model training engine are activated to generate patient historical cycle feature correlation data and an initial personalized ovulation induction strategy recommendation model, including: Based on the preset feature screening rules, the multi-period clinical data in the initial patient dataset are processed. The SHAP method is used to screen key predictive factors and quantify the contribution of each feature, generating a core predictive factor set and feature importance weight information. Based on the model training requirements, the dataset partitioning design of the core predictive factor set is carried out. Following the sample size standard of EPV≥20, the partitioning ratio and screening rules of the training set and test set are determined, and a dataset partitioning scheme is generated. Combining the sample imbalance condition and the model generalization requirements, the pre-stored SMOTE sample balancing optimization algorithm and model evaluation rules are called to perform oversampling on the training set to generate balanced training samples. The core predictive factor set, dataset partitioning scheme, and balanced training samples are integrated, and the sliding window time series processing module and multi-algorithm model training engine are launched to generate patient historical cycle feature correlation data and initial individualized ovulation induction strategy recommendation model.

3. The method as described in claim 2, characterized in that, The initial individualized ovulation induction strategy recommendation model is trained using balanced training samples and a test set. The training results are then combined with model evaluation rules to generate a target ovulation induction prediction model, including: The balanced training samples and test sets were split into protocol training data subsets and dosage training data subsets according to the prediction target. These subsets were then input into the protocol prediction module and dosage preliminary recommendation module of the initial individualized ovulation induction strategy recommendation model, respectively. Iterative training was carried out simultaneously to generate model training process data for each module. The training process incorporated sliding window temporal modeling logic. The protocol prediction module used multi-cycle correlation data, including the patient's previous COS protocol effects and ovarian response, to optimize classification parameters. The dosage prediction module used multi-cycle correlation data, including previous Gn dosage and number of oocytes retrieved, to optimize regression parameters. Furthermore, the dosage prediction module split the data subsets according to the first to fourth cycles and subsequent cycles, and independently trained the sub-model for each cycle. Based on the preset model evaluation rules, the training process data of the two modules are quantitatively evaluated, and a model performance evaluation report containing the two modules and the sub-models of each cycle is generated. The evaluation covers internal five-fold cross-validation and external independent dataset validation. The accuracy rate is the core evaluation indicator for the scheme prediction module, and the R² value is the core evaluation indicator for the dose prediction module. Based on the evaluation results, the model algorithm of the two modules is optimized by hierarchical adaptation. The LightGBM algorithm is selected for both the protocol prediction module and the dose prediction module in the first / second COS cycle, and the ExtraTrees algorithm is selected for both the protocol prediction module and the dose prediction module in the third / fourth and subsequent cycles. Based on the evaluation performance of the sub-model in each cycle, the hyperparameters of the algorithm are further fine-tuned to generate the optimized parameter configuration of the two modules and the sub-model in each cycle. By integrating and optimizing the dual-module parameter configuration, the parameter configuration of each cycle sub-model, and the performance evaluation results, the collaborative linkage logic between the COS protocol prediction module and the multi-cycle gonadotropin dosage prediction module is clarified, a complete model architecture is constructed, and a target ovulation induction prediction model is generated.

4. The method as described in claim 1, characterized in that, Based on a hierarchical classification strategy and weighted ensemble rules, the evaluation results of the balanced training samples and test set are processed to generate COS scheme classification prediction signals and multi-cycle Gn dose prediction adjustment information, including: Based on the hierarchical classification strategy, the balanced training samples are processed in layers to generate candidate signals for COS scheme classification. The hierarchical classification first trains a binary classification model for the COS scheme category with the largest sample size and adjusts the decision probability threshold, and then trains a multi-class classification model for the remaining categories. Combined with SMOTE oversampling, the sample imbalance is solved. The test set evaluation results and classification candidate signals are fused and optimized using a weighted integration rule to generate the COS scheme classification prediction signal. The weighted integration assigns differentiated weights to the binary classifier prediction results corresponding to different categories. The evaluation results of the balanced training samples and the test set are split according to the COS periodic stratification rule to generate periodic-specific training data; A multi-cycle Gn dose prediction sub-model is constructed based on cycle-specific training data. The model parameters are optimized by combining the differences in clinical characteristics of each cycle, and multi-cycle Gn dose prediction adjustment information is generated.

5. The method as described in claim 1, characterized in that, The COS protocol classification prediction signal, multi-cycle Gn dose prediction adjustment information, and clinical applicability validation data are processed to generate individualized protocol matching scores and dose recommendation accuracy calibration information, including: We retrieved COS protocol classification prediction signals, multi-cycle Gn dose prediction and adjustment information, and clinical applicability verification data to establish a three-dimensional evaluation dimension. The clinical applicability verification data included real clinical data such as the previous COS protocol implementation effect, ovarian response feedback, and the occurrence of complications. Based on the preset protocol adaptation threshold and dose-response matching benchmark, the predicted signals, dose adjustment information and clinical validation data are compared to identify the correction signals for protocol adaptation deviation and dose-response mismatch. Protocol adaptation deviation is determined by whether the degree of fit between the predicted protocol and the clinically optimal protocol meets the standard, and dose-response mismatch is determined by whether the number of oocytes retrieved corresponding to the predicted dose and the ovarian response meet the expected benchmark. Collect protocol matching data and dose calibration cases in different clinical scenarios, construct a protocol-dose-clinical effect correlation model, calculate the model's fit coefficient at the current prediction parameter node, and use this coefficient to generate individualized protocol matching score and dose recommendation accuracy calibration information.

6. The method as described in claim 5, characterized in that, Based on the target ovulation induction prediction model, the user input data, individualized protocol matching score, and dose recommendation accuracy calibration information that can be used for prediction are processed to generate and output individualized COS protocols and Gn starting dose recommendations, including: Extract the core prediction logic of the target ovulation induction prediction model, and collect user input data that can be used for prediction, individualized protocol matching score, and dose recommendation accuracy calibration information. Establish the feature-score-calibration correlation mapping relationship among the three to generate user feature-fit score-dose calibration correlation data. Based on user characteristics, fit scores, and dose calibration correlation data, and targeting the clinical individual characteristics of the target user, the deviation between the matching score and the optimal fit threshold is calculated in real time. Combined with dose calibration information, individualized adjustment parameters are generated. These individualized adjustment parameters include the optimization direction of the COS protocol type, the adjustment range of the initial Gn dose, and the applicable cycle. The parameter generation is synchronized with the update frequency of the user's multi-cycle clinical data. Based on the user's COS treatment cycle stratification, and combined with the ovarian response characteristics and model algorithm adaptation differences of each cycle, the recommendation results for each cycle are periodically adapted and adjusted to generate cycle-specific recommendation strategies. Among them, the cycle-specific recommendation strategies clarify the optimal COS regimen, Gn starting dose range and clinical monitoring focus for each cycle.

7. A device for predicting individualized ovulation induction protocols and gonadotropin initiation doses based on multi-cycle time-series modeling, characterized in that, The device is configured to execute the method for predicting the individualized ovulation induction protocol and the initial dose of gonadotropins based on multi-cycle time-series modeling as described in any one of claims 1 to 6 by executing the executable instructions.

8. An electronic device, characterized in that, include: First processor; and memory for storing executable instructions of the first processor; The first processor is configured to execute the method for predicting the individualized ovulation induction protocol and the initial dose of gonadotropins based on multi-cycle time-series modeling as described in any one of claims 1 to 6 by executing the executable instructions.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the second processor, it implements the individualized ovulation induction protocol and the method for predicting the initial dose of gonadotropins based on multi-cycle time-series modeling as described in any one of claims 1 to 6.

Images