A Machine Learning-Based Method and System for Recommending Initial Gonadotropin Dosage
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-05-28
- Publication Date
- 2026-08-14
AI Technical Summary
[0004]但是，现有技术中还没有通过对大数据的分析，开发机器学习算法，通过研究不同患者在不同促排卵药物剂量下不同反应的病例，获得更加科学的药物剂量优化策略，以起到辅助临床决策的技术方案
[0023]本发明的于机器学习的初始促性腺激素用量推荐方法及系统，可以精准预测初始促性腺激素用量供医生参考，从而显著提高IVF-ET治疗的成功率和安全性，基于9200条临床数据训练，预测准确率达到84%，能够有效降低卵巢过度刺激综合征(OHSS)发生风险，实现真正的个体化用药推荐。
Smart Images
Figure CN122575619A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of artificial intelligence technology, specifically relating to a method and system for recommending initial gonadotropin dosage based on machine learning. Background Technology
[0002] In some medical settings, gonadotropins are required. For example, to increase the success rate of in vitro fertilization (IVF), a certain amount of ovulation-inducing drugs is injected to increase the number of released eggs and improve the conception rate. In this case, the initial dosage of ovulation-inducing drugs must be carefully controlled. Too little dosage will not produce enough eggs, while too much dosage can damage the woman's ovaries and even cause ovarian hyperstimulation syndrome (OHSS). Currently, most doctors rely on clinical experience to determine the appropriate initial dosage of ovulation-inducing drugs for patients. This method is overly dependent on the doctor's qualifications. By deeply analyzing existing clinical treatment data, artificial intelligence algorithms can be used to establish an evaluation system to find the most suitable initial dosage of drugs, thereby complementing the doctor's treatment plan.
[0003] In existing technologies, machine learning algorithms are used to attempt to predict the outcome of assisted reproductive technology (ART). For example, ovarian response is predicted using the baseline characteristics of IVF patients (including age, BMI, smoking status, and initial estrogen levels). Oocyte prediction is performed on multiple subjects using a multi-layered and linear neural network with a certain training-test ratio.
[0004] However, current technologies do not yet offer a solution for developing machine learning algorithms based on big data analysis, studying different patients' responses to different ovulation-inducing drug dosages, and obtaining more scientific drug dosage optimization strategies to assist clinical decision-making. Summary of the Invention
[0005] To overcome the shortcomings of existing technologies and achieve the goal of recommending the optimal initial dosage of gonadotropins, this invention adopts the following technical solution:
[0006] A machine learning-based method for recommending initial gonadotropin dosages includes the following steps:
[0007] Collect historical data on gonadotropin users, including physiological information, initial gonadotropin dosage, and total number of eggs retrieved.
[0008] A predictive model for the number of ovulations is constructed and trained, with the physiological information data and the initial gonadotropin dosage as model inputs and the final total number of ovulations as model output.
[0009] Based on different initial gonadotropin dosage levels and corresponding physiological information data of the users to be treated with gonadotropins, the total number of eggs retrieved at different dosage levels is predicted using the trained prediction model.
[0010] Furthermore, the physiological information data includes age, height, weight, anti-Müllerian hormone levels, basal follicle-stimulating hormone levels, basal luteinizing hormone levels, basal estradiol levels, and total number of antral follicles.
[0011] Furthermore, a nonlinear coupling feature is constructed based on the synergistic relationship between the anti-Müllerian hormone and the total number of antral follicles to characterize ovarian reserve. By obtaining the ratio and product features of anti-Müllerian hormone and the total number of antral follicles, the synergistic effect of reserve function that cannot be reflected by a single indicator can be captured.
[0012] Furthermore, the historical data is preprocessed to remove missing values and / or outliers, and then normalized to unify data at different scales into the same range, so as to adapt to the machine learning model and facilitate model training.
[0013] Furthermore, the prediction model adopts the extreme gradient boosting model, which includes a set of sequentially connected decision trees for preliminary prediction, residual correction, and combined optimization. Based on the historical data, a training set and a test set are constructed. The extreme gradient boosting model is trained through the training set, and the difference between the predicted number of ovulations and the actual number of ovulations at each level is statistically analyzed through the test set.
[0014] A machine learning-based initial gonadotropin dosage recommendation system, including a historical data acquisition module and an ovulation number prediction module;
[0015] The historical data acquisition module is used to collect historical data of gonadotropin users. The historical data includes physiological information data, initial gonadotropin dosage, and final total number of eggs retrieved.
[0016] The ovulation count prediction module is used to construct and train a prediction model for predicting the number of ovulations. The physiological information data and the initial gonadotropin dosage are used as model inputs, and the final total number of eggs retrieved is used as model output. Based on different initial gonadotropin dosage levels and the corresponding physiological information data of the gonadotropin user, the trained prediction model predicts the final total number of eggs retrieved at different levels.
[0017] Furthermore, in the historical data acquisition module, physiological information data includes age, height, weight, anti-Müllerian hormone levels, basal follicle-stimulating hormone levels, basal luteinizing hormone levels, basal estradiol levels, and total number of antral follicles.
[0018] Furthermore, in the historical data acquisition module, a nonlinear coupling feature is constructed. Based on the synergistic relationship between the anti-Müllerian hormone and the total number of antral follicles, the ratio and product features of the anti-Müllerian hormone and the total number of antral follicles are obtained to capture the synergistic effect of reserve function that cannot be reflected by a single indicator.
[0019] Furthermore, in the historical data acquisition module, the historical data is preprocessed to remove missing values and / or outliers, and then normalized to unify data of different scales into the same range, so as to adapt to the machine learning model and facilitate model training.
[0020] Furthermore, in the ovulation number prediction module, the prediction model adopts the extreme gradient boosting model, which includes a set of decision trees connected in sequence, for preliminary prediction, residual correction and combined optimization.
[0021] Furthermore, the ovulation number prediction module, based on the training set and test set constructed from the historical data, trains the extreme gradient boosting model through the training set and statistically analyzes the difference between the predicted ovulation number and the actual ovulation number at each level through the test set.
[0022] The advantages and beneficial effects of this invention are as follows:
[0023] The initial gonadotropin dosage recommendation method and system based on machine learning of the present invention can accurately predict the initial gonadotropin dosage for doctors' reference, thereby significantly improving the success rate and safety of IVF-ET treatment. Based on training with 9,200 clinical data, the prediction accuracy reaches 84%, which can effectively reduce the risk of ovarian hyperstimulation syndrome (OHSS) and achieve truly personalized medication recommendation. Attached Figure Description
[0024] Figure 1 This is a flowchart of the method in an embodiment of the present invention.
[0025] Figure 2 This is a structural architecture diagram of the machine learning model in an embodiment of the present invention.
[0026] Figure 3 This is a performance diagram of the model in an embodiment of the present invention.
[0027] Figure 4 This is a schematic diagram of the reasoning process for different levels of Gn usage in an embodiment of the present invention.
[0028] Figure 5 This is a schematic diagram of the system structure in an embodiment of the present invention. Detailed Implementation
[0029] The specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings. It should be understood that the specific embodiments described herein are for illustration and explanation only and are not intended to limit the present invention.
[0030] like Figure 1 As shown, the machine learning-based method for recommending initial gonadotropin (Gn) dosages utilizes accumulated patient medical record data to train a model that depicts the effect of specific initial Gn dosages on individual patients. The trained model is then deployed on the doctor's computer to recommend appropriate Gn dosages, eliminating the need for doctors to rely solely on personal experience. The collected historical patient information includes age, height, weight, hormone levels, Gn dosage, and the total number of eggs retrieved. Age is used as input to the machine learning model, and the total number of eggs retrieved is used as output. When using this model, various individual patient characteristics and different Gn dosage levels are input, and the model ultimately predicts the total number of eggs retrieved at different Gn dosages, providing doctors with relevant reference information.
[0031] The method of the present invention specifically includes the following steps:
[0032] Step S1: Collect a dataset of features of female patients undergoing in vitro fertilization (IVF) to construct historical patient data features. This dataset includes various basic physiological information indicators of female patients during actual treatment, information on the doctor's initial dosage of gonadotropin (Gn), and information on the total number of eggs retrieved by the patient. This information is divided into two parts: the first part consists of basic physiological information indicators and information on the doctor's initial dosage of gonadotropin (Gn), which serve as input features for the machine learning model; the second part consists of information on the total number of eggs retrieved by the patient, which serves as output features for the machine learning model.
[0033] In this embodiment of the invention, 9200 patient samples were collected, and the input features of the machine learning model included:
[0034] 1) Patient's age;
[0035] 2) Patient's height;
[0036] 3) Patient weight;
[0037] 4) Patient's anti-Müllerian hormone level (AMH, used to evaluate ovarian reserve function and determine the number of antral oocytes that may form in the ovary; the higher the AMH value, the greater the oocyte reserve; the lower the value, the worse the ovarian function).
[0038] 5) Patient's baseline follicle-stimulating hormone (FSH) level (hereinafter referred to as b-FSH, which promotes follicle development and maturation).
[0039] 6) Patient's baseline luteinizing hormone (LH, hereinafter referred to as b-LH, is a hormone that, together with FSH, promotes follicle maturation).
[0040] 7) The patient's baseline estradiol level (estradiol, hereinafter referred to as β-E2, is mainly produced by theca cells before ovulation and has positive and negative feedback regulatory effects on the hypothalamus and pituitary gland).
[0041] 8) Total number of antral follicles in the patient (Total number of antral follicles = number of left antral follicles + number of right antral follicles, hereinafter referred to as AFC);
[0042] 9) Initial dosage of Gn for patients (This refers to the first dose of Gn given to the patient during the treatment cycle, which is a key factor for doctors to control the total number of ovulations in patients).
[0043] In this embodiment of the invention, the output features of the machine learning model include:
[0044] 1) The final total number of eggs retrieved (8-14 is considered an ideal number for IVF by the doctor).
[0045] Step S2: After collecting the patient's historical data features, the data features need to be preprocessed. The purpose of data preprocessing is roughly divided into two parts: removing abnormal data and data normalization.
[0046] The purpose of removing outlier data is to prevent some patients' indicators from being significantly higher than normal, which could cause the training model to lose its fit to this part of the data and ultimately lead to a decline in model performance.
[0047] After removing all missing or outlier values from the data collected in step S1, the data needs to be normalized because the normal ranges of different indicators vary. The purpose of data normalization is to unify the patient's information at different scales into the 0-1 range to adapt to the machine learning model and facilitate model training.
[0048] In this embodiment of the invention, the minimum-maximum normalization method is used, and its formula is as follows:
[0049]
[0050] in, Represents the original data. express The minimum value among the features of the sample data. express The maximum value among the features of the sample data. This represents the new data after the normalization operation.
[0051] Furthermore, this invention also constructs nonlinear coupling features. There is a synergistic relationship between anti-Müllerian hormone (AMH) and antral follicle count (AFC) in characterizing ovarian reserve. This invention calculates the ratio feature R_AMA = AMH / AFC and the product feature P_AMA = AMH × AFC to capture the synergistic effect of reserve function that cannot be reflected by a single indicator.
[0052] Step S3: Build and train a machine learning model to predict the number of ovulations in patients.
[0053] After data preprocessing, a series of patient historical characteristic data were obtained, which constituted the training set for the machine learning model. Subsequently, the training set was input into the machine learning model for training. Specifically, the basic physiological information data indicators mentioned in step S1 and preprocessed in step S2, along with the doctor's initial gonadotropin (Gn) dosage information, were used as the input to the machine learning model, denoted as x; the total number of oocytes retrieved by the patient, mentioned in step S1 and preprocessed in step S2, was used as the predicted output of the machine learning model, denoted as y. The prediction model is generated and trained.
[0054] In this embodiment of the invention, the Extreme Gradient Boosting Model is employed, such as... Figure 2 The image shows a limit gradient boosting model containing 1000 decision trees. As shown in step S1, the input feature layer has a dimension of 10 (clinical input features of 10 patients), the output prediction layer has a dimension of 1 (outputting the predicted total number of oocytes retrieved), and each decision tree has a maximum depth of 6 layers. The learning rate is 0.005. The reason for using a model with a large number of parameters is:
[0055] 1) A larger training set can significantly improve the model's generalization ability; 9,200 data points cover a wide range of clinical feature combinations, effectively reducing the risk of overfitting and ensuring that the model can work stably and reliably in real clinical environments, which is also the key to the success of medical AI applications.
[0056] 2) Large datasets significantly improve prediction accuracy; ample data allows the extreme gradient boosting model to learn more accurately the complex nonlinear relationship between 10 clinical features and the total number of oocytes retrieved, providing clinicians with reliable medication recommendations.
[0057] 3) Large-scale training data enhances the reliability of medical decisions; 9200 data points meet the statistical significance requirements of evidence-based medicine, significantly improving clinicians' trust in AI recommendations, ensuring the scientific validity and safety of Gn dosage recommendations, and providing strong support for assisted reproductive treatment. In this embodiment of the invention, the final accuracy rate on the test set was 84%, and the mean absolute error between the model's predicted number of oocytes and the actual number of oocytes retrieved was 1.74.
[0058] Furthermore, on the test set, the percentage of samples showing the difference between the predicted and actual ovulation numbers for each grade was calculated, such as... Figure 3 As shown, the final result is as follows:
[0059] 1) In 42% of the samples, the absolute error in predicting the number of ovulations was within 1.
[0060] 2) 27% of the samples had an absolute error in predicting the number of ovulations between 1 and 2.
[0061] 3) 15% of the samples had an absolute error of 2-3 in predicting the number of ovulations;
[0062] 4) 16% of the samples had an absolute error of more than 3 in predicting the number of ovulations.
[0063] Step S4: Save the ovulation number prediction machine learning model obtained from training, and input the gonadotropin dosage at different levels along with the patient's characteristic data into the machine learning model to infer the ovulation number.
[0064] Specifically, after obtaining the trained model, the different levels of Gn dosage, along with the patient's basic information data features, are input into the trained machine learning model for inference. After obtaining the predicted number of ovulations under different Gn dosage levels, the appropriate Gn dosage level for the number of ovulations is selected and recommended to the doctor as a reference dosage. The dosage level can be set manually, and the doctor usually sets a reasonable Gn dosage range of 75-300 IU.
[0065] like Figure 4 As shown in this embodiment of the invention, the dosage of Gn is divided into 21 levels, that is, every 12.5 points from 50 to 300 is one level. Then, the dosage of Gn divided into levels is also normalized. The operation process is the same as in step S2. Finally, the model outputs the predicted ovulation data information under the 21 dosage levels of Gn. Select one or more data points that are closest to 8-14 and use their corresponding Gn dosage level as the recommended Gn dosage value for doctors to refer to.
[0066] Step S5: Evaluate the accuracy of the model. Re-collect 790 samples and input them into the model for validation. The results are as follows:
[0067] There are 300 cases with an error range of 0-1 in the number of eggs retrieved.
[0068] There were 202 cases with an error of 1-2 in the number of eggs retrieved.
[0069] There were 139 samples with an error of 2-3 in the number of eggs retrieved.
[0070] There were 149 cases where the error in the number of eggs retrieved was greater than 3.
[0071] The assessment accuracy rate was 81.14%.
[0072] like Figure 5 As shown, the initial gonadotropin dosage recommendation system based on machine learning includes a historical data acquisition module and an ovulation number prediction module.
[0073] The historical data acquisition module is used to collect historical data of gonadotropin users. Historical data includes physiological information data, initial gonadotropin dosage, and the final total number of oocytes retrieved. Physiological information data includes age, height, weight, anti-Müllerian hormone level, basal follicle-stimulating hormone level, basal luteinizing hormone level, basal estradiol level, and total number of antral follicles. In addition, the historical data needs to be preprocessed to remove missing values and / or outliers, and then normalized to unify data at different scales into the same interval to adapt to the machine learning model and facilitate model training.
[0074] The ovulation count prediction module is used to construct and train a prediction model for ovulation count. It takes the physiological information data and the initial gonadotropin dosage as model inputs and the final total number of retrieved eggs as model output. Based on different initial gonadotropin dosage levels and the corresponding physiological information data of the user, the trained prediction model predicts the final total number of retrieved eggs at different levels. Specifically, the prediction model employs a limit gradient boosting model, which includes a set of sequentially connected decision trees for preliminary prediction, residual correction, and combined optimization. Furthermore, a training set and a test set are constructed based on historical data. The limit gradient boosting model is trained using the training set, and the difference between the predicted ovulation count and the actual ovulation count at each level is statistically analyzed using the test set.
[0075] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features therein. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.
Claims
1. A method for recommending initial gonadotropin dosage based on machine learning, characterized in that... Includes the following steps: Collect historical data on gonadotropin users, including physiological information, initial gonadotropin dosage, and total number of eggs retrieved. A predictive model for the number of ovulations is constructed and trained, with the physiological information data and the initial gonadotropin dosage as model inputs and the final total number of ovulations as model output. Based on different initial gonadotropin dosage levels and corresponding physiological information data of the users to be treated with gonadotropins, the total number of eggs retrieved at different dosage levels is predicted using the trained prediction model.
2. The method for recommending initial gonadotropin dosage based on machine learning according to claim 1, characterized in that: The physiological information data includes age, height, weight, anti-Müllerian hormone levels, basal follicle-stimulating hormone levels, basal luteinizing hormone levels, basal estradiol levels, and total number of antral follicles.
3. The method for recommending initial gonadotropin dosage based on machine learning according to claim 2, characterized in that: A nonlinear coupling feature is constructed based on the synergistic relationship between the anti-Müllerian hormone and the total number of antral follicles to characterize ovarian reserve. By obtaining the ratio and product features of anti-Müllerian hormone and the total number of antral follicles, the synergistic effect of reserve function that cannot be reflected by a single indicator can be captured.
4. The method for recommending initial gonadotropin dosage based on machine learning according to claim 1, characterized in that: The historical data is preprocessed to remove missing values and / or outliers, and then normalized to unify data from different scales into the same range.
5. The method for recommending initial gonadotropin dosage based on machine learning according to claim 1, characterized in that: The prediction model adopts the extreme gradient boosting model, which includes a set of sequentially connected decision trees for preliminary prediction, residual correction and combined optimization. The training set and test set are constructed from the historical data. The extreme gradient boosting model is trained through the training set, and the difference between the predicted number of ovulations and the actual number of ovulations at each level is statistically analyzed through the test set.
6. A machine learning-based initial gonadotropin dosage recommendation system, comprising a historical data acquisition module and an ovulation count prediction module, characterized in that: The historical data acquisition module is used to collect historical data of gonadotropin users. The historical data includes physiological information data, initial gonadotropin dosage, and final total number of eggs retrieved. The ovulation count prediction module is used to construct and train a prediction model for predicting the number of ovulations. The physiological information data and the initial gonadotropin dosage are used as model inputs, and the final total number of eggs retrieved is used as model output. Based on different initial gonadotropin dosage levels and the corresponding physiological information data of the gonadotropin user, the trained prediction model predicts the final total number of eggs retrieved at different levels.
7. The initial gonadotropin dosage recommendation system based on machine learning according to claim 6, characterized in that: The historical data acquisition module includes physiological information data such as age, height, weight, anti-Müllerian hormone levels, basal follicle-stimulating hormone levels, basal luteinizing hormone levels, basal estradiol levels, and total number of antral follicles.
8. The method for recommending initial gonadotropin dosage based on machine learning according to claim 7, characterized in that: In the historical data acquisition module, a nonlinear coupling feature is constructed. Based on the synergistic relationship between the anti-Müllerian hormone and the total number of antral follicles, the ratio and product features of the anti-Müllerian hormone and the total number of antral follicles are obtained to capture the synergistic effect of reserve function that cannot be reflected by a single indicator.
9. The initial gonadotropin dosage recommendation system based on machine learning according to claim 6, characterized in that: In the historical data acquisition module, the historical data is preprocessed to remove missing values and / or outliers, and then normalized to unify data of different scales into the same range.
10. The machine learning-based initial gonadotropin dosage recommendation system according to claim 6, characterized in that: In the ovulation number prediction module, the prediction model adopts the extreme gradient boosting model, which includes a set of sequentially connected decision trees for preliminary prediction, residual correction and combined optimization. Based on the historical data, a training set and a test set are constructed. The extreme gradient boosting model is trained through the training set, and the difference between the predicted ovulation number and the actual ovulation number at each level is statistically analyzed through the test set.