Auxiliary reproduction multi-node clinical decision-making method based on machine learning

Abstract

The invention relates to the technical field of application of artificial intelligence in assisted reproduction technology, and discloses an assisted reproduction multi-node clinical decision-making method based on machine learning. The method comprises the following steps: collecting sample data, obtaining sample features, and carrying out conversion and interpolation on the sample features; performing predictive variable screening on the sample features by using a Spearman correlation coefficient; dividing the data of the complete sample into a training set, a test set and a verification set according to a sample proportion of 8: 1: 1, taking a predicted variable obtained by screening as an independent variable in the training set, and taking accumulated live birth within 2 years after single egg taking as a dependent variable; respectively constructing prediction models in different decisions of four stages of a controlled ovarian stimulation scheme, a gonadotropin initiation amount and the like, and correcting to obtain prediction values; and carrying out hyper-parameter adjustment by using the test set, and comparing live birth outcomes of the crowds which accord with and do not accord with the recommendation by using the verification set. According to the method, corresponding cumulative live yield prediction can be provided for various feasible schemes, and a single optimal path is not recommended.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence in assisted reproductive technology, specifically to a multi-node clinical decision-making method for assisted reproduction based on machine learning. Background Technology

[0002] The incidence of infertility is increasing year by year, gradually becoming a public health issue that cannot be ignored. Assisted reproductive technologies, especially in vitro fertilization-embryo transfer (IVF-ET), have received much attention in recent years as a key measure for treating infertility. With the development of embryo vitrification and freezing-thawing technology, the cumulative live birth rate achieved through multiple embryo transfers after a single oocyte retrieval is gaining increasing clinical acceptance: in recent years, several high-level randomized controlled clinical trials have used the cumulative live birth rate as the primary outcome measure.

[0003] IVF-ET is a sequential treatment highly dependent on precise clinical decision-making. All patients must undergo the following steps in sequence: 1. Selection of a controlled ovarian stimulation protocol, i.e., choosing from long, short, and antagonist protocols; 2. Selection of the initial gonadotropin (PG) dose, commonly ranging from 75-150 IU and 175-200 IU; 3. Selection of the human chorionic gonadotropin (hCG) dose, commonly ranging from <6000 IU to 6000 IU; 4. The decision of whether or not to perform fresh embryo transfer. Decisions at any stage must be based on a clear understanding of the patient's current stage. For example, the hCG dose selection must be made after the controlled ovarian stimulation protocol and PG initial dose are determined, and the resulting impact on follicular development must also be considered. Due to this sequential nature, choices at each stage can only be made sequentially; it is neither possible nor necessary to coordinate all decisions beforehand.

[0004] Clinical guidelines provide guidance on the selection of controlled ovarian stimulation protocols, but the complexity of patient conditions and responses during assisted reproductive processes often necessitates clinicians developing personalized assisted reproductive measures. This leads to significant differences in the success rate of assisted reproduction for the same patient, due to variations in the experience of those developing the protocols. The development of artificial intelligence supports non-empirical personalized IVF-ET, and the development of big data-based assisted reproductive systems is an inevitable choice in addressing infertility.

[0005] This application proposes a machine learning-based multi-node clinical decision-making method for assisted reproduction to achieve personalized assisted reproduction based on artificial intelligence. Summary of the Invention

[0006] The purpose of this invention is to provide a multi-node clinical decision-making method for assisted reproduction based on machine learning, so as to solve the problems in the prior art mentioned in the background section.

[0007] To achieve the above objectives, the present invention provides the following technical solution: a multi-node clinical decision-making method for assisted reproduction based on machine learning. The method includes the following steps: Step 1: Collect sample data and obtain sample features; Step 1.1: Set sample data collection criteria. Patients meeting the criteria were included in the follow-up: The patients were women who underwent in vitro fertilization-embryo transfer at multiple national or provincial reproductive centers, including the Reproductive Hospital Affiliated to Shandong University, between January 2015 and January 2023 due to infertility caused by various reasons such as fallopian tube obstruction or adhesion. Their residences were spread across various cities in my country. After excluding those with chromosomal abnormalities, a total of 73,413 patients were included in the analysis. Two years after follow-up, the final outcome was obtained: cumulative live births from a single oocyte retrieval. Patient age, body mass index, ovarian reserve indicators, ovarian response to drug stimulation, oocyte and embryo laboratory indicators, embryo transfer, and live birth information were all simultaneously recorded in the electronic medical record system during treatment and retrieved from various reproductive centers with the patients' informed consent. During retrieval, the patients' medical record numbers and other personally identifiable information were removed to ensure desensitization security. Data transfer was completed by designated personnel using the hospital's intranet system, and the time frame and purpose of use were strictly registered and reported, prohibiting use in external network environments. Data use was supervised by the hospital's ethics committee, and the entire process complied with the relevant provisions of the "Information Security Technology: Guidelines for Health and Medical Data Security." Considering the risk of re-identification, the desensitized data of the large population was not publicly disclosed to fully protect patient privacy. The desensitized data of two patients were specially provided with their informed consent so that those skilled in the art could use the model provided by this patent to implement corresponding instances. Step 1.2: Select sample features related to pregnancy outcomes from the sample data according to preset criteria: Spearman analysis confirmed that the features were associated with cumulative live births, i.e., the correlation coefficient was greater than 0.05 and the statistical significance p value was less than 0.05; Step 2: Transform and interpolate the sample features described in Step 1.2. The specific transformation and interpolation process is as follows: Step 2.1: Transform and encode the sample features: Standardize and normalize the basic data such as female age, anti-Müllerian hormone (AMH), antral follicle count (AFC), and basal follicle-stimulating hormone (bFSH), as well as the laboratory data such as the number of follicles ≥1.4cm on HCG day, the number of oocytes retrieved, and the number of normally fertilized embryos, so that they meet the requirements of the prediction model input. Step 2.2: If some features in the sample have missing values, multiple imputation is used to fill in the missing values ​​for different features: the missing value ratio of all features is less than 5%; multiple imputation generates five complete datasets, and each missing value is imputed using methods such as predicted mean matching to maintain the distribution of the original data and the relationship between variables. The complete dataset that is closest to the statistical features of the original sample is selected for subsequent steps; if there are no missing values, no imputation is needed, and a complete sample without missing values ​​is finally formed and proceeds to step 3; Step 3: Using the Spearman correlation coefficient, the complete sample formed in Step 2.2 is used to screen candidate predictive variables. Specifically, indicators with an absolute value of Spearman correlation coefficient greater than 0.05 and a significance value (p value) less than 0.05 with cumulative live births are selected and included in the predictive model in Step 4 as needed, along with the controlled ovarian stimulation protocol, gonadotropin (Gn) initiation amount, human chorionic gonadotropin (HCG) dosage, and whether fresh embryo transfer occurs on day 3. Step 4: Divide the complete sample data into training, test, and validation sets in an 8:1:1 ratio; construct a cumulative live birth prediction model based on a generalized additive model using the training set samples formed in steps 2, 3, and 4: use the processed data as independent variables and the cumulative live birth rate within 2 years after a single oocyte retrieval as the dependent variable to construct the model. The general formula of the model is:

[0008] Provide cumulative live birth predictions for each patient; Before selecting a controlled ovarian stimulation protocol, the specific formula in the model predicting cumulative live births based on age, anti-Müllerian hormone (AMH), antral follicle count (AFC), basal follicle-stimulating hormone (bFSH), and body mass index (BMI) is as follows: g(y)=[f1(x1)+ℇ1]+[f2(x2)+ℇ2]+[f3(x3)+ℇ3]+[f4(x4)+ℇ4]+[f5(x5)+ℇ5]; Where y represents cumulative live births, x1, x2, x3, x4, and x5 are age, anti-Müllerian hormone, antral follicle count, basal follicle-stimulating hormone, and body mass index after normalization and standardization, respectively, and f1(x1)+ℇ1, f2(x2)+ℇ2, f3(x3)+ℇ3, f4(x4)+ℇ4, and f5(x5)+ℇ5 are the smoothing functions and error terms used to fit the five parameters, respectively. These are automatically selected by the prediction model based on restricted maximum likelihood estimation to accurately reflect the true changes in the data; the test set is used to adjust the hyperparameters, and the validation set is used to evaluate the model accuracy. Step 5: Perform Platt correction on the predicted values ​​output in Step 4 to improve the reliability and interpretability of the model output probabilities; That is, in the test set, the predicted values ​​s of each sample are used. i A logistic regression model is constructed based on whether or not there is accumulated live births. The specific formula is as follows:

[0009] Where s i P represents the initial prediction value of the cumulative live birth prediction model. i The corrected predicted values ​​are represented by A, where A is the slope and B is the bias, derived from the linear fit between the predicted cumulative live birth probability and the actual cumulative live birth in the test set: probability exponent y. i = ;P i Sigmoidization of the probability exponent ( That is, to convert it into a probability between 0 and 1, so as to accurately reflect the possibility of live birth; Step 6: Using the generalized additive model and Platt correction model constructed in Steps 4 and 5, calculate the corrected predicted values ​​in the validation set and evaluate their accuracy in predicting cumulative live births. Step 7: Based on the action selection stage, select variables from Step 3 to predict cumulative live births under different actions and construct predictive models. The multiple models simulate the multiple decision-making processes in clinical practice where clinical information is gradually obtained: for example, after determining the protocol, further decide on the gonadotropin initiation dosage; after obtaining the estradiol level and the number of follicles reaching 1.4 cm in diameter on the trigger day, decide on the human chorionic gonadotropin dosage; after obtaining information such as the number of retrieved oocytes, the number of oocytes that have released the first polar body, the number of normally fertilized embryos, and the number of high-quality embryos on day 3, decide whether to perform fresh embryo transfer on day 3. The selection of predictive and outcome variables in each model construction is as follows, all conforming to the reference information and specific decisions in clinical decision-making: Step 7.1, Selection of controlled ovarian stimulation protocol: female age, anti-Müllerian hormone, antral follicle count, basal follicle-stimulating hormone, body mass index, and whether or not primary infertility exists; generalized additive model and logistic regression model were trained and validated in populations using long protocol, short protocol, antagonist protocol, ultra-long protocol, and other protocols for controlled ovarian stimulation. Step 7.2, Selection of gonadotropin initiation dose (Gn): female age, anti-Müllerian hormone, antral follicle count, basal follicle-stimulating hormone, body mass index, primary infertility status, and controlled ovarian stimulation protocol; the model was constructed and validated in populations using Gn initiation at 75-150 IU, 175-200 IU, 225-250 IU, 275-300 IU, and 325-450 IU respectively; Step 7.3, Selection of Human Chorionic Gonadotropin (hCG) Dosage: Female age, anti-Müllerian hormone, antral follicle count, basal follicle-stimulating hormone (FSH) level, body mass index, primary infertility status, controlled ovarian stimulation protocol, gonadotropin initiation and total dose, estradiol level on trigger day, and number of follicles with a diameter of 1.4 cm; Models were constructed and validated in populations using <6000 IU, 6000 IU, 8000 IU, and 10000~14000 IU hCG respectively; Step 7.4, Selection of Day 3 Fresh Embryo Transfer: Female age, anti-Müllerian hormone, antral follicle count, basal follicle-stimulating hormone (FSH), body mass index, controlled ovarian stimulation protocol, gonadotropin initiation dose and total dose, estradiol level on trigger day, number of follicles with diameter reaching 1.4 cm, number of oocytes retrieved, number of oocytes that released the first polar body, number of normally fertilized embryos, and number of high-quality embryos on Day 3; The model was constructed and validated in populations that underwent Day 3 fresh embryo transfer and those that did not; Step 8: Using the model obtained from Steps 4, 5, 6, and 7, calculate the corrected predicted values ​​for the entire included population and identify those with poor intervention choices: When choosing between controlled ovarian stimulation protocols, gonadotropin initiation doses, human chorionic gonadotropin (hCG) dosages, and fresh embryo transfer on day 3, if the intervention with the highest expected live birth probability increases the expected live birth probability by 5% compared to the actual intervention used, and the 95% confidence intervals of the two interventions do not overlap, then the patient is considered to have made a poor intervention choice at this stage: 5% represents a clinically perceptible improvement; a difference of less than 5% is a minor improvement and has little clinical significance; the 95% confidence interval represents the range of uncertainty in the estimated results, and no overlap in the 95% confidence intervals of the two interventions means that the difference is statistically reliable, i.e., compared to the optimal intervention recommended by the model, the intervention actually chosen by the patient caused a significant decrease in the cumulative live birth rate.

[0010] Compared with the prior art, the beneficial effects of the present invention are: 1. At different key decision points in assisted reproductive treatment, this invention can provide cumulative live birth rate predictions for multiple feasible options, rather than recommending a single best path. This helps doctors and patients make more personalized choices based on clinical conditions and personal preferences with full knowledge.

[0011] 2. For unsuccessful IVF-ET assisted reproductive cycles, this invention can provide data-driven references for analyzing the reasons for failure; by revealing the potential correlation between different protocols and the final outcome, it can help protocol developers deepen their understanding, optimize decision-making, and thus accelerate their clinical experience learning and knowledge iteration. Attached Figure Description

[0012] Figure 1 This is a flowchart of a multi-node clinical decision-making method for assisted reproduction based on machine learning, according to the present invention. Figure 2 This is a schematic diagram of the Spearman correlation coefficients between various data points and cumulative live births in a machine learning-based multi-node clinical decision-making method for assisted reproduction according to the present invention. Figure 3 This is a schematic diagram illustrating the changes in data specificity and sensitivity at each stage in an embodiment of a machine learning-based multi-node clinical decision-making method for assisted reproduction according to the present invention. Figure 4 This is a schematic diagram comparing laboratory indicators of women with and without good treatment protocols in an embodiment of a multi-node clinical decision-making method for assisted reproduction based on machine learning according to the present invention. The following parameters were compared sequentially in the groups with and without good treatment protocols: estradiol level on trigger day (a), number of follicles 1.4 cm and above (b), number of retrieved oocytes (c), number of oocytes released from the first polar body (d), number of normally fertilized embryos (e), number of high-quality embryos on day 3 (f), number of usable blastocysts (g), and number of high-quality blastocysts (h). The **** above the bars indicates that the differences in the corresponding indicators between the two groups (good and bad treatment protocols) are highly significant, with a statistical significance p-value of less than 0.0001. Figure 5 This is a schematic diagram comparing the cumulative live births of women with good and bad treatment plans in an embodiment of a multi-node clinical decision-making method for assisted reproduction based on machine learning according to the present invention; the **** above the bar chart indicates that the difference in the cumulative live birth rate between the two groups (good and bad treatment plans) is very significant, with a statistical significance p value of less than 0.0001; Figure 6 This is a schematic diagram of a preferred embodiment of a machine learning-based multi-node clinical decision-making method for assisted reproduction according to the present invention. The upper part of the diagram shows the baseline data and actual IVF-ET decision of a patient, and the lower part shows the cumulative live birth rate predicted for various decisions at each stage. The black box marks the decision with the highest cumulative live birth probability, which is the recommended decision. Figure 7 This is a schematic diagram of an unsatisfactory embodiment of a multi-node clinical decision-making method for assisted reproduction based on machine learning according to the present invention; the upper part of the figure shows the baseline data and actual IVF-ET decision of a patient, and the lower part shows the cumulative live birth rate predicted for various decisions at each stage. The black box marks the decision with the highest cumulative live birth probability, which is the recommended decision. Detailed Implementation

[0013] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Example

[0014] Please see Figures 1-7 The present invention provides a technical solution: A machine learning-based multi-node clinical decision-making method for assisted reproduction, comprising the following steps: Step 1: Collect sample data and obtain sample features; Step 1.1: Set sample data collection criteria. Patients meeting the criteria were included in the follow-up: The patients were women who underwent in vitro fertilization-embryo transfer at multiple national or provincial reproductive centers, including the Reproductive Hospital Affiliated to Shandong University, between January 2015 and January 2023 due to infertility caused by various reasons such as fallopian tube obstruction or adhesion. Their residences were spread across various cities in my country. After excluding those with chromosomal abnormalities, a total of 73,413 patients were included in the analysis. Two years after follow-up, the final outcome was obtained: cumulative live births from a single oocyte retrieval. Patient age, body mass index, ovarian reserve indicators, ovarian response to drug stimulation, oocyte and embryo laboratory indicators, embryo transfer, and live birth information were all simultaneously recorded in the electronic medical record system during treatment and retrieved from various reproductive centers with the patients' informed consent. During retrieval, the patients' medical record numbers and other personally identifiable information were removed to ensure desensitization security. Data transfer was completed by designated personnel using the hospital's intranet system, and the time frame and purpose of use were strictly registered and reported, prohibiting use in external network environments. Data use was supervised by the hospital's ethics committee, and the entire process complied with the relevant provisions of the "Information Security Technology: Guidelines for Health and Medical Data Security." Considering the risk of re-identification, the desensitized data of the large population was not publicly disclosed to fully protect patient privacy. The desensitized data of two patients were specially provided with their informed consent so that those skilled in the art could use the model provided by this patent to implement corresponding instances. Step 1.2: Select sample features related to pregnancy outcomes from the sample data according to preset criteria: Spearman analysis confirmed that the features were associated with cumulative live births, i.e., the correlation coefficient was greater than 0.05 and the statistical significance p value was less than 0.05; Step 2: Transform and interpolate the sample features described in Step 1.2. The specific transformation and interpolation process is as follows: Step 2.1: Transform and encode the sample features: Standardize and normalize the basic data such as female age, anti-Müllerian hormone (AMH), antral follicle count (AFC), and basal follicle-stimulating hormone (bFSH), as well as the laboratory data such as the number of follicles ≥1.4cm on HCG day, the number of oocytes retrieved, and the number of normally fertilized embryos, so that they meet the requirements of the prediction model input. Step 2.2: If some features in the sample have missing values, multiple imputation is used to fill in the missing values ​​for different features: the missing value ratio of all features is less than 5%; multiple imputation generates five complete datasets, and each missing value is imputed using methods such as predicted mean matching to maintain the distribution of the original data and the relationship between variables. The complete dataset that is closest to the statistical features of the original sample is selected for subsequent steps; if there are no missing values, no imputation is needed, and a complete sample without missing values ​​is finally formed and proceeds to step 3; Step 3: Using the Spearman correlation coefficient, the complete sample formed in Step 2.2 is used to screen candidate predictive variables. Specifically, indicators with an absolute value of Spearman correlation coefficient greater than 0.05 and a significance value (p value) less than 0.05 with cumulative live births are selected and included in the predictive model in Step 4 as needed, along with the controlled ovarian stimulation protocol, gonadotropin (Gn) initiation amount, human chorionic gonadotropin (HCG) dosage, and whether fresh embryo transfer occurs on day 3. Step 4: Divide the complete sample data into training, test, and validation sets in an 8:1:1 ratio; construct a cumulative live birth prediction model based on a generalized additive model using the training set samples formed in steps 2, 3, and 4: use the processed data as independent variables and the cumulative live birth rate within 2 years after a single oocyte retrieval as the dependent variable to construct the model. The general formula of the model is:

[0015] Provide cumulative live birth predictions for each patient; Before selecting a controlled ovarian stimulation protocol, the specific formula in the model predicting cumulative live births based on age, anti-Müllerian hormone (AMH), antral follicle count (AFC), basal follicle-stimulating hormone (bFSH), and body mass index (BMI) is as follows: g(y)=[f1(x1)+ℇ1]+[f2(x2)+ℇ2]+[f3(x3)+ℇ3]+[f4(x4)+ℇ4]+[f5(x5)+ℇ5]; Where y represents cumulative live births, x1, x2, x3, x4, and x5 are age, anti-Müllerian hormone, antral follicle count, basal follicle-stimulating hormone, and body mass index after normalization and standardization, respectively, and f1(x1)+ℇ1, f2(x2)+ℇ2, f3(x3)+ℇ3, f4(x4)+ℇ4, and f5(x5)+ℇ5 are the smoothing functions and error terms used to fit the five parameters, respectively. These are automatically selected by the prediction model based on restricted maximum likelihood estimation to accurately reflect the true changes in the data; the test set is used to adjust the hyperparameters, and the validation set is used to evaluate the model accuracy. Step 5: Perform Platt correction on the predicted values ​​output in Step 4 to improve the reliability and interpretability of the model output probabilities; That is, in the test set, the predicted values ​​s of each sample are used. i A logistic regression model is constructed based on whether or not there is accumulated live births. The specific formula is as follows:

[0016] Where s i P represents the initial prediction value of the cumulative live birth prediction model. i The corrected predicted values ​​are represented by A, where A is the slope and B is the bias, derived from the linear fit between the predicted cumulative live birth probability and the actual cumulative live birth in the test set: probability exponent y. i = ;P i Sigmoidization of the probability exponent ( That is, to convert it into a probability between 0 and 1, so as to accurately reflect the possibility of live birth; Step 6: Using the generalized additive model and Platt correction model constructed in Steps 4 and 5, calculate the corrected predicted values ​​in the validation set and evaluate their accuracy in predicting cumulative live births. Step 7: Based on the action selection stage, select variables from Step 3 to predict cumulative live births under different actions and construct predictive models. The multiple models simulate the multiple decision-making processes in clinical practice where clinical information is gradually obtained: for example, after determining the protocol, further decide on the gonadotropin initiation dosage; after obtaining the estradiol level and the number of follicles reaching 1.4 cm in diameter on the trigger day, decide on the human chorionic gonadotropin dosage; after obtaining information such as the number of retrieved oocytes, the number of oocytes that have released the first polar body, the number of normally fertilized embryos, and the number of high-quality embryos on day 3, decide whether to perform fresh embryo transfer on day 3. The selection of predictive and outcome variables in each model construction is as follows, all conforming to the reference information and specific decisions in clinical decision-making: Step 7.1, Selection of controlled ovarian stimulation protocol: female age, anti-Müllerian hormone, antral follicle count, basal follicle-stimulating hormone, body mass index, and whether or not primary infertility exists; generalized additive model and logistic regression model were trained and validated in populations using long protocol, short protocol, antagonist protocol, ultra-long protocol, and other protocols for controlled ovarian stimulation. Step 7.2, Selection of gonadotropin initiation dose (Gn): female age, anti-Müllerian hormone, antral follicle count, basal follicle-stimulating hormone, body mass index, primary infertility status, and controlled ovarian stimulation protocol; the model was constructed and validated in populations using Gn initiation at 75-150 IU, 175-200 IU, 225-250 IU, 275-300 IU, and 325-450 IU respectively; Step 7.3, Selection of Human Chorionic Gonadotropin (hCG) Dosage: Female age, anti-Müllerian hormone, antral follicle count, basal follicle-stimulating hormone (FSH) level, body mass index, primary infertility status, controlled ovarian stimulation protocol, gonadotropin initiation and total dose, estradiol level on trigger day, and number of follicles with a diameter of 1.4 cm; Models were constructed and validated in populations using <6000 IU, 6000 IU, 8000 IU, and 10000~14000 IU hCG respectively; Step 7.4, Selection of Day 3 Fresh Embryo Transfer: Female age, anti-Müllerian hormone, antral follicle count, basal follicle-stimulating hormone (FSH), body mass index, controlled ovarian stimulation protocol, gonadotropin initiation dose and total dose, estradiol level on trigger day, number of follicles with diameter reaching 1.4 cm, number of oocytes retrieved, number of oocytes that released the first polar body, number of normally fertilized embryos, and number of high-quality embryos on Day 3; The model was constructed and validated in populations that underwent Day 3 fresh embryo transfer and those that did not; Step 8: Using the model obtained from Steps 4, 5, 6, and 7, calculate the corrected predicted values ​​for the entire included population and identify those with poor intervention choices: When choosing between controlled ovarian stimulation protocols, gonadotropin initiation doses, human chorionic gonadotropin (hCG) dosages, and fresh embryo transfer on day 3, if the intervention with the highest expected live birth probability increases the expected live birth probability by 5% compared to the actual intervention used, and the 95% confidence intervals of the two interventions do not overlap, then the patient is considered to have made a poor intervention choice at this stage: 5% represents a clinically perceptible improvement; a difference of less than 5% is a minor improvement and has little clinical significance; the 95% confidence interval represents the range of uncertainty in the estimated results, and no overlap in the 95% confidence intervals of the two interventions means that the difference is statistically reliable, i.e., compared to the optimal intervention recommended by the model, the intervention actually chosen by the patient caused a significant decrease in the cumulative live birth rate.

[0017] In this embodiment, the population that met the model's recommendations at each stage and the population that did not meet the model's recommendations at any stage were screened. After matching data on female age, anti-Müllerian hormone, antral follicle count, basal follicle-stimulating hormone, and body mass index, the population that met the model's recommendations at each stage had more follicles reaching 1.4cm on the trigger day, more oocytes retrieved and released from the first polar body, more normally fertilized embryos and more high-quality embryos on day 3, and significantly improved cumulative live births. The cumulative live birth rate of the best protocol reached 70%, while the cumulative live birth rate of the poor protocol was 53.9%, and the significance values ​​of both were less than 0.0001, indicating that following the model's recommended protocol can improve controlled ovarian stimulation response, laboratory data, and pregnancy outcomes. In this example, the results show a 36-year-old woman whose controlled ovarian stimulation protocol, Gn initiation amount and HCG dosage all met the model recommendations, and who did not transfer fresh embryos on day 3 as required by the model. She ultimately achieved a live birth and had 7 usable blastocysts cryopreserved. It should be noted that: when analyzing failed assisted reproductive cycles, for example, an analysis of a 30-year-old woman with excellent ovarian reserve but obesity, whose expected cumulative live birth probability was over 80%, but who ultimately did not achieve a live birth; Based on the situation, the model recommends an ultra-long ovulation induction scheme, suggesting the use of 250 IU Ugn to start and 6000 IU HCG; The patient actually used an antagonist protocol for ovulation induction, using 150 IU U n and 8000 IU HCG to initiate the process, resulting in only 6 oocytes retrieved (expected 12-20). It should be noted that, based on this learning of the model, it is pointed out that obese patients with excessive ovarian reserve should not blindly use antagonist regimens and low initiation doses to prevent overstimulation of the ovaries. Instead, the insensitivity of obese patients to Gn should be fully considered, and other regimens and higher initiation doses should be actively tried.

[0018] It should be noted that the predictive model was built and validated in a large population, including elderly patients, patients with recurrent implantation failure, and patients with chronic diseases such as diabetes and hypertension. This example is only for illustration and does not mean that the model is only applicable to people with excellent ovarian reserve but who are obese.

Claims

1. A multi-node clinical decision-making method for assisted reproduction based on machine learning, characterized in that: The method includes the following steps: Step 1: Collect sample data and obtain sample features; Step 2: Transform and interpolate the sample features described in Step 1; Step 3: Use Spearman correlation coefficient to screen candidate predictive variables for the complete sample formed in Step 2. That is, select indicators with an absolute value of Spearman correlation coefficient greater than 0.05 and a significance value of less than 0.05 with the cumulative live birth. These indicators, along with the controlled ovarian stimulation protocol, gonadotropin initiation dose, human chorionic gonadotropin dosage, and whether fresh embryo transfer is performed on day 3, are included in the predictive model in Step 4 as needed. Step 4: Divide the complete sample data into training, test, and validation sets in an 8:1:1 ratio; construct a cumulative live birth prediction model based on a generalized additive model using the training set samples formed in steps 2, 3, and 4: use the processed data as independent variables and the cumulative live births within 2 years after a single oocyte retrieval as the dependent variable to construct the model; use the test set for the model hyperparameter K value, i.e., the maximum degrees of freedom of the independent variables; when the K value is 6, the model prediction accuracy reaches its maximum value; Step 5: Perform Platt correction on the predicted values ​​output in Step 4 to improve the reliability and interpretability of the model output probabilities; Step 6: Using the generalized additive model and Platt correction model constructed in Steps 4 and 5, calculate the corrected predicted values ​​in the validation set and evaluate their accuracy in predicting cumulative live births. Step 7: Based on the action selection stage, select from the variables in Step 3 to predict the cumulative live births under different actions and construct a predictive model; the multiple models constructed simulate the multiple decision-making process of gradually obtaining clinical information in clinical practice: such as further deciding the gonadotropin initiation dose after determining the plan; deciding the human chorionic gonadotropin dosage after obtaining the estradiol level and the number of follicles with a diameter of 1.4 cm on the trigger day; and deciding whether to perform fresh embryo transfer on day 3 after obtaining information such as the number of oocytes retrieved, the number of oocytes that have released the first polar body, the number of normally fertilized embryos, and the number of high-quality embryos on day 3. Step 8: Using the model obtained from Steps 4, 5, 6, and 7, calculate the corrected predicted values ​​for the entire included population and identify the population with poor intervention choices: When choosing between controlled ovarian stimulation protocols, gonadotropin initiation doses, human chorionic gonadotropin (hCG) dosages, and fresh embryo transfer on day 3, if the intervention with the highest expected live birth probability increases the expected live birth probability by 5% compared to the actual intervention used, and the 95% confidence intervals of the two do not overlap, then the patient is considered to have poor intervention choices at this stage.

2. The multi-node clinical decision-making method for assisted reproduction based on machine learning according to claim 1, characterized in that: Step 1 specifically includes: Step 1.1: Set sample data collection criteria, and patients who meet the criteria are included in follow-up: Obtain the final outcome: cumulative live births from a single oocyte retrieval; Step 1.2: Select sample features related to pregnancy outcomes from the sample data according to the preset criteria: Spearman analysis confirmed that the features were associated with cumulative live births, that is, the correlation coefficient was greater than 0.05 and the statistical significance p value was less than 0.

05.

3. The multi-node clinical decision-making method for assisted reproduction based on machine learning according to claim 1, characterized in that: In step 2, the specific transformation and interpolation process of sample features is as follows: Step 2.1: Transform and encode the sample features: Standardize and normalize the obtained data on female age, anti-Müllerian hormone, antral follicle count, basal follicle-stimulating hormone, number of follicles ≥1.4cm on HCG day, number of oocytes retrieved, and number of normally fertilized embryos to make them meet the requirements of the prediction model input. Step 2.2: If some features in the sample have missing values, multiple imputation is used to fill in the missing values ​​of different features: the missing value ratio of all features is less than 5%; multiple imputation generates five complete datasets, and each missing value is imputed by methods such as predicted mean matching to maintain the distribution of the original data and the relationship between variables. The complete dataset that is closest to the statistical features of the original sample is selected to proceed to the next step; if there are no missing values, no imputation is required, and a complete sample without missing values ​​is finally formed and proceeds to step 3.

4. The multi-node clinical decision-making method for assisted reproduction based on machine learning according to claim 1, characterized in that: Step 7 specifically includes: Step 7.1, Selection of controlled ovarian stimulation protocol: female age, anti-Müllerian hormone, antral follicle count, basal follicle-stimulating hormone, body mass index, and whether or not primary infertility exists; generalized additive model and logistic regression model were trained and validated in populations using long protocol, short protocol, antagonist protocol, and ultra-long protocol controlled ovarian stimulation respectively. Step 7.2, Selection of gonadotropin initiation dose: female age, anti-Müllerian hormone, antral follicle count, basal follicle-stimulating hormone, body mass index, primary infertility status, and controlled ovarian stimulation protocol; the model was constructed and validated in populations using 75-150 IU, 175-200 IU, 225-250 IU, 275-300 IU, and 325-450 IU Gn initiation. Step 7.3, Selection of Human Chorionic Gonadotropin (hCG) Dosage: Female age, anti-Müllerian hormone, antral follicle count, basal follicle-stimulating hormone (FSH) level, body mass index, primary infertility status, controlled ovarian stimulation protocol, gonadotropin initiation and total dose, estradiol level on trigger day, and number of follicles with a diameter of 1.4 cm; Models were constructed and validated in populations using <6000 IU, 6000 IU, 8000 IU, and 10000~14000 IU hCG respectively; Step 7.4, Selection of Day 3 Fresh Embryo Transfer: Female age, anti-Müllerian hormone, antral follicle count, basal follicle-stimulating hormone, body mass index, controlled ovarian stimulation protocol, gonadotropin initiation dose and total dose, estradiol level on trigger day, number of follicles with diameter reaching 1.4 cm, number of oocytes retrieved, number of oocytes that have released the first polar body, number of normally fertilized embryos, and number of high-quality embryos on Day 3; The model was constructed and validated in populations that underwent Day 3 fresh embryo transfer and those that did not.

5. The multi-node clinical decision-making method for assisted reproduction based on machine learning according to claim 1, characterized in that: In step 4, the formula for the generalized additive model is: Provide cumulative live birth predictions for each patient; Before selecting a controlled ovarian stimulation protocol, the specific formula in the model predicting cumulative live births based on age, anti-Müllerian hormone, antral follicle count, basal follicle-stimulating hormone, and body mass index is as follows: g(y)=[f1(x1)+ℇ1]+[f2(x2)+ℇ2]+[f3(x3)+ℇ3]+[f4(x4)+ℇ4]+[f5(x5)+ℇ5]; Where y represents the cumulative live births, x1, x2, x3, x4, and x5 are the normalized and standardized ages, anti-Müllerian hormone, antral follicle count, basal follicle-stimulating hormone, and body mass index, respectively, and f1(x1)+ℇ1, f2(x2)+ℇ2, f3(x3)+ℇ3, f4(x4)+ℇ4, and f5(x5)+ℇ5 are the smoothing functions and error terms used to fit the five parameters, respectively, which are automatically selected by the prediction model based on restricted maximum likelihood estimation to accurately reflect the true changes in the data; The test set is used to tune hyperparameters, and the validation set is used to evaluate model accuracy.

6. The multi-node clinical decision-making method for assisted reproduction based on machine learning according to claim 1, characterized in that: Step 5 specifically involves, in the test set, using the predicted values ​​s of each sample... i A logistic regression model is constructed based on whether or not there is accumulated live births. The specific formula is as follows: Where s i P represents the initial prediction value of the cumulative live birth prediction model. i The corrected predicted values ​​are represented by A, where A is the slope and B is the bias, derived from the linear fit between the predicted cumulative live birth probability and the actual cumulative live birth in the test set: probability exponent y. i = ;P i Sigmoidization of the probability exponent ( That is, to convert it into a probability between 0 and 1, so as to accurately reflect the likelihood of a live birth.

Images