A method and system for predicting outcomes of preimplantation genetic testing

Abstract

The application provides a method and system for predicting the outcome of preimplantation genetic testing, relating to the technical field of intelligent medical treatment, which comprises: acquiring multi-source characteristics of embryos; inputting the multi-source characteristics into an outcome prediction model to obtain three-classification prediction probabilities and prediction uncertainty; performing closed-loop optimization on the outcome prediction model based on an offline reinforcement learning mechanism, including acquiring feedback samples, outputting parameter update actions from a strategy network, training the strategy network in an offline playback environment based on a multi-objective reward function that takes into account prediction accuracy, probability calibration and update stability, and controlling the model update amplitude through a joint constraint mechanism to obtain updated model parameters. The application builds a closed-loop mechanism for prediction, feedback and offline safe optimization, enabling the prediction model to continuously absorb new real-world data and continuously and stably evolve under the premise of ensuring medical safety, significantly improving the continuous optimization capability and safety of the model in long-term clinical application.

Description

Technical Field

[0001] This invention relates to the field of intelligent medical technology, and in particular to a method and system for predicting the outcome of preimplantation genetic testing. Background Technology

[0002] Preimplantation genetic testing (PGT) is an important technique in the field of assisted reproduction used to assess the genetic status of embryos before embryo transfer. It involves biopsy sampling and genetic analysis during the in vitro embryo culture stage to determine whether the embryo has chromosomal abnormalities or specific pathogenic gene mutations, thereby providing a basis for clinical transplantation decisions.

[0003] Existing technologies use artificial intelligence models to predict embryo implantation potential, pregnancy outcome, or euploidy probability by analyzing embryo morphological images, time-lapse imaging sequences, or combining a small number of clinical parameters. However, most predictive models are static, meaning they are trained once on historical samples and then permanently deployed. The model parameters are not updated with the new real-world test results and follow-up information generated during clinical application. This static characteristic makes it difficult for the model to adapt to changes in clinical sample distribution, evolution of testing conditions, and differences in case structure at different times, making it difficult to guarantee predictive stability and generalization ability in long-term use. If conventional online incremental training methods are used for updates, the lack of safety constraints may lead to model performance fluctuations or high-risk misjudgments, which greatly limits the reliability of predictive models in medical scenarios requiring high safety. Summary of the Invention

[0004] To address the aforementioned technical problems, this invention provides a method and system for predicting the outcome of preimplantation genetic testing.

[0005] The first aspect of the present invention provides a method for predicting the outcome of preimplantation genetic testing, comprising: obtaining multi-source characteristics of the embryo, wherein the multi-source characteristics include genetic disease characteristics, parental clinical characteristics, ovulation induction cycle characteristics, and embryonic development characteristics; The multi-source features are input into the outcome prediction model to obtain the three-class prediction probability and prediction uncertainty. The outcome prediction model is optimized using an offline reinforcement learning mechanism. This optimization includes: acquiring feedback samples, which are associated with the multi-source features, the three-class prediction probability, the prediction uncertainty, and the target outcome label; updating the action by the policy network based on the output parameters of the current state, where the current state is jointly represented by the latent representation of the multi-source features, the three-class prediction probability, and the prediction uncertainty; training the policy network in an offline playback environment based on a multi-objective reward function, which balances prediction accuracy, probability calibration, and update stability, and controls the model update magnitude through a joint constraint mechanism; and obtaining the updated outcome prediction model parameters based on the trained policy network.

[0006] Furthermore, the multi-source features are obtained in the following manner: Collect genetic disease characteristics including inheritance pattern, pathogenic locus type, and paternal or maternal origin information; Clinical characteristics of the parents were collected, including maternal age, anti-Müllerian hormone level, number of antral follicles, body mass index, paternal age, and semen parameters. Collect information on ovulation induction cycle characteristics, including ovulation induction protocol, starting dose, total dose, number of days of ovulation induction, and number of oocytes retrieved; The data collected included embryonic development and grading characteristics such as fertilization method, cleavage stage development status, blastocyst grade, culture days, and embryo morphology scoring. The continuous variables in the genetic disease characteristics, parental clinical characteristics, ovulation induction cycle characteristics, and embryonic development and grading characteristics are standardized, the categorical variables are encoded, and missing values ​​are filled in, and then combined to form the multi-source feature vector.

[0007] Furthermore, the prediction uncertainty is calculated based on the entropy uncertainty of the three-class prediction results, and the calculation formula is as follows: ; in, This represents the predicted probability for the corresponding category.

[0008] Furthermore, the multi-objective reward function is expressed as: ; in, For accuracy-related benefits, To calibrate the payoff term for probability, To conservatively update the constraint terms, , These are the weighting coefficients.

[0009] Furthermore, the accuracy gain term is defined as: ; in, For predicting categories, For the real category, This is the penalty coefficient for misclassification; The probability calibration benefit term is defined as: ; in, For the model to the true category The predicted probability; The conservative update constraint term is defined as follows: ; in, Indicates time Model parameters; Indicates the current optimization strategy; The divergence is Kullback-Leibler.

[0010] Furthermore, the joint constraint mechanism is implemented by minimizing the total loss function, which is: ; in, Losses due to three categories of supervision, For balance coefficient, To reinforce the loss of learning value; The reinforcement learning value loss is obtained by minimizing the Bellman residual of the state-action value function, expressed as: ; in, For the value network to be trained, For the target network, This is the discount factor.

[0011] A second aspect of the present invention provides a preimplantation genetic testing outcome prediction system, comprising: a feature acquisition module for acquiring multi-source features of the embryo, wherein the multi-source features include genetic disease features, parental clinical features, ovulation induction cycle features, and embryonic development features; The outcome prediction module is used to input the multi-source features into the outcome prediction model to obtain the three-class prediction probability and prediction uncertainty. The model optimization module is used to perform closed-loop optimization of the outcome prediction model based on an offline reinforcement learning mechanism. The closed-loop optimization includes: acquiring feedback samples, which are associated with the multi-source features, the three-class prediction probabilities, the prediction uncertainty, and the target outcome label; updating the action by the policy network outputting parameters according to the current state, where the current state is jointly represented by the latent representation of the multi-source features, the three-class prediction probabilities, and the prediction uncertainty; training the policy network based on a multi-objective reward function in an offline playback environment, where the multi-objective reward function takes into account prediction accuracy, probability calibration, and update stability, and controls the model update magnitude through a joint constraint mechanism; and obtaining the updated outcome prediction model parameters based on the trained policy network.

[0012] A third aspect of the present invention provides an electronic device including a memory, a processor, and a program stored in the memory and running on the processor, wherein the processor executes the program to implement the steps in the preimplantation genetic testing outcome prediction method as described in the first aspect of the present invention.

[0013] A fourth aspect of the present invention provides a computer-readable storage medium having a program stored thereon that, when executed by a processor, implements the steps in the preimplantation genetic testing outcome prediction method as described in the first aspect of the present invention.

[0014] A fifth aspect of the present invention provides a computer program product comprising software code, wherein a program in the software code performs steps in the preimplantation genetic testing outcome prediction method as described in the first aspect of the present invention.

[0015] Compared with existing technologies, the present invention provides a method and system for predicting the outcome of preimplantation genetic testing. By constructing feedback samples that link predicted information with actual outcomes, and in an offline playback environment (rather than a real online environment) composed of historical feedback data, the method uses a multi-objective reward function and joint constraint mechanism to control the model update magnitude. This allows the model to absorb new sample information to improve prediction performance while avoiding performance fluctuations and safety risks caused by online trial and error updates, thereby improving the model's continuous learning ability and update controllability in this medical scenario. Attached Figure Description

[0016] The accompanying drawings, which form part of this disclosure, are used to provide a further understanding of this disclosure. The illustrative embodiments of this disclosure and their descriptions are used to explain this disclosure and do not constitute an undue limitation of this disclosure.

[0017] Figure 1 This is a flowchart of the preimplantation genetic testing outcome prediction method provided in Embodiment 1 of the present invention.

[0018] Figure 2The SHAP importance bar chart for the first 20 key features provided in Embodiment 1 of the present invention.

[0019] Figure 3 This is the confusion matrix of the XGBoost model provided in Embodiment 1 of the present invention during time verification.

[0020] Figure 4 The ROC curve of the XGBoost model provided in Embodiment 1 of the present invention during time verification.

[0021] Figure 5 This is a waterfall plot of a transplantable embryo in the local model interpretation based on the SHAP method provided in Embodiment 1 of the present invention.

[0022] Figure 6 The waterfall plot is a local model interpretation based on the SHAP method provided in Embodiment 1 of the present invention, showing a single embryo requiring genetic counseling.

[0023] Figure 7 This is a waterfall plot of a non-transferable embryo in the local model interpretation based on the SHAP method provided in Embodiment 1 of the present invention.

[0024] Figure 8 This is an architecture diagram of the preimplantation genetic testing outcome prediction system provided in Embodiment 2 of the present invention. Detailed Implementation

[0025] It should be noted that the following detailed descriptions are exemplary and intended to provide further illustration of the invention. Unless otherwise specified, all technical and scientific terms used in this invention have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains.

[0026] It should be noted that the terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit the scope of exemplary embodiments according to the invention. As used herein, unless the context clearly indicates otherwise, the singular form is intended to include the plural form as well. Furthermore, it should be understood that the terms “comprising” and “having”, and any variations thereof, are intended to cover non-exclusive inclusion, for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0027] Where there is no conflict, the embodiments and features in the embodiments of the present invention can be combined with each other.

[0028] All data acquisition in this embodiment is carried out in accordance with laws and regulations and with user consent, and the data is used legally.

[0029] Example 1 like Figure 1 This embodiment provides a method for predicting the outcome of preimplantation genetic testing, including: S1. Obtain multi-source characteristics of the embryo, including genetic disease characteristics, parental clinical characteristics, ovulation induction cycle characteristics, and embryonic development characteristics.

[0030] First, data collection and sample definition are carried out. Clinical cycle data of those undergoing preimplantation genetic testing are collected, and cycle samples with complete follow-up information and traceable key variables are selected to form an embryo-level dataset. Each embryo-level sample includes at least the following information: (1) genetic disease-related information, including inheritance mode, genetic source, pathogenic locus category, and related chromosome information; (2) parental clinical information, including maternal age, anti-Müllerian hormone level, number of antral follicles, body mass index, history of adverse pregnancy and childbirth, paternal age, and semen parameters; (3) ovulation induction cycle information, including ovulation induction protocol, starting dose, total dose, number of days of ovulation induction, and number of oocytes retrieved; (4) embryo development information, including fertilization method, cleavage stage development status, blastocyst grade, culture days, and embryo morphology score; (5) preimplantation genetic testing true outcome label. The true outcome label is divided into three categories: transplantable, requiring genetic counseling, and not transplantable. The sample size is not less than 1500 embryos. To ensure the effectiveness of subsequent feedback optimization, we will also record the actual test results, clinical treatment outcomes, and necessary follow-up results.

[0031] In one specific embodiment, based on the data collection method described above, embryo-level data from 1737 subjects undergoing preimplantation genetic testing were included. As shown in Tables 1 and 2, embryo samples were divided into transferable embryos (n=452), embryos requiring genetic counseling (n=408), and non-transferable embryos (n=877) according to the true outcome label. The distribution differences among the groups in multiple dimensions, including female baseline characteristics, male baseline characteristics, controlled ovulation induction cycle characteristics, and embryonic development characteristics, were systematically compared. Statistical results showed significant differences (P<0.05) among the three groups in several key characteristics, including maternal age, paternal age, baseline antral follicle count, sperm concentration and motility, initial gonadotropin dose, pathogenic gene inheritance pattern, blastocyst grade, and culture days. This provides a real and reliable clinical data foundation for subsequent multi-source feature construction and model training.

[0032] Table 1. Baseline data of parental lineage and ovulation induction cycle characteristics of 1737 subjects who underwent preimplantation genetic testing.

[0033] Table 2. Baseline data of genetic and embryonic developmental characteristics of 1737 subjects who underwent preimplantation genetic testing.

[0034] Note: Adverse pregnancy history includes recurrent implantation failure, recurrent miscarriage, and history of induced labor due to genetic factors. Values ​​are expressed as mean ± standard error, median (P25, P75), or percentage of cases. A p-value < 0.05 was considered statistically significant.

[0035] Based on this, multi-source features are constructed. Based on the raw data collected above, a multi-source feature vector for prediction is constructed. The feature vector contains 15 to 40 features, specifically including: (1) genetic disease features, which may include discrete variables such as dominant inheritance, recessive inheritance, paternal origin, maternal origin, autosomal correlation, and sex chromosome correlation; (2) parental clinical features, which may include continuous features and categorical features; (3) ovulation induction cycle features; (4) embryonic development and grading features, which may include developmental status and blastocyst grade information at stages D1, D3, D5, or D6. A unified coding rule is used for the same thing to form a structured input vector. .

[0036] In one specific embodiment, the multi-source features are obtained in the following ways: collecting genetic disease features including inheritance pattern, pathogenic locus type, and paternal or maternal origin information; collecting parental clinical features including maternal age, anti-Müllerian hormone level, number of antral follicles, body mass index, paternal age, and semen parameters; collecting ovulation induction cycle features including ovulation induction protocol, starting dose, total dose, number of days of ovulation induction, and number of oocytes retrieved; and collecting embryonic development and grading features including fertilization method, cleavage stage development status, blastocyst grade, culture days, and embryo morphology score.

[0037] Then, the obtained multi-source feature vectors are preprocessed. The preprocessing includes: (1) standardizing continuous variables by Z-score standardization or min-max normalization; (2) encoding categorical variables by one-hot encoding, sequential encoding or target encoding; (3) imputing missing values ​​by K-nearest neighbor imputation or multiple imputation, where the K value of K-nearest neighbor is 3 to 10 and the number of multiple imputation iterations is 5 to 20; (4) identifying and processing outliers by box plot thresholding, quantile truncation or clinically reasonable interval truncation; (5) dividing the data into training set, validation set and test set, either by time stratification or by subject level, to avoid leakage of different embryos from the same subject into different data subsets at the same time. All preprocessing parameters are fitted only on the training set and applied to the validation set and test set to prevent information leakage. The continuous variables in the genetic disease characteristics, parental clinical characteristics, ovulation induction cycle characteristics, and embryonic development and grading characteristics are standardized, the categorical variables are encoded, and missing values ​​are filled in, and then combined to form the multi-source feature vector.

[0038] Among the aforementioned multi-source features, genetic disease features reflect the potential genetic risk background of the embryo; parental clinical features reflect the physiological state of the mother and father; ovulation induction cycle features reflect the drug response and ovarian reserve in the current treatment cycle; and embryonic development features reflect the growth quality and morphological performance of the embryo during in vitro culture. By uniformly acquiring and constructing structured feature vectors from these four categories of multi-dimensional information routinely obtainable in clinical settings—ranging from macroscopic parental status to microscopic embryonic development—data from different sources, scales, and types can be jointly modeled in the same mathematical space. This provides the prediction model with information-rich and formatted input, solving the technical problem of directly utilizing multi-source heterogeneous medical data. Compared to existing methods that rely solely on a single embryo image or a single morphological score, the multi-source feature system constructed in this step can more comprehensively cover the complex factors affecting the outcome of embryonic genetic testing, thereby improving the comprehensiveness and specificity of the prediction.

[0039] S2. Input the multi-source features into the outcome prediction model to obtain the three-class prediction probability and prediction uncertainty.

[0040] Specifically, the preprocessed feature vectors are input into the outcome prediction model to establish a three-class prediction model for preimplantation genetic testing outcomes. The outcome prediction model is an ensemble learning model, including one or more of the following: adaptive boosting model, random forest model, gradient boosting tree model, and extreme gradient boosting model. Alternatively, it can be replaced with a logistic regression model, lightweight gradient boosting model, neural network model, etc., as long as the model can output three-class probability. The model outputs a three-class probability vector: ; in, , , Let represent the probabilities that the target embryo is predicted to be transplantable, require genetic counseling, and is not transplantable, respectively, and satisfy the following conditions: .

[0041] Meanwhile, to enhance the model's sensitivity to high-risk misclassification samples, this invention introduces an uncertainty characterization mechanism. The prediction uncertainty is calculated based on the entropy uncertainty of the three-class classification prediction results, using the following formula: ; in, This represents the predicted probability for the corresponding category. Represents any category in the three-category label set. ,in, This indicates portability. The representative needs genetic counseling. This uncertainty is not portable. When the probability distribution is relatively uniform, the uncertainty is high; when the probability distribution is sharp, the uncertainty is low. The uncertainty calculation method can also be replaced by model ensemble variance, Monte Carlo Dropout, etc., in alternative solutions. This uncertainty index enables the system to proactively identify high-risk samples with ambiguous prediction boundaries. In subsequent feedback optimization stages, this uncertainty serves as a component of the reinforcement learning state, allowing the policy network to prioritize samples in high-uncertainty regions, improving the targeting and efficiency of updates.

[0042] The outcome prediction model does not output a simple binary classification result, but rather a probability distribution and prediction uncertainty encompassing three categories: transplantable, requiring genetic counseling, and non-transplantable. The category requiring genetic counseling is modeled and output separately, enabling the system to distinguish samples in the gray zone between clearly usable and clearly unusable, allowing clinicians to further assess the results. This improves the discriminative power and clinical usability of the prediction results. Gray zone samples include embryos with a low proportion of chimeras, uncertain outcomes, or those requiring further manual interpretation. The output prediction uncertainty provides a crucial risk perception signal for subsequent closed-loop optimization. The model exhibits higher uncertainty for samples it is unsure about, providing a basis for triggering safety strategies.

[0043] In one specific embodiment, the multi-source feature vectors of the aforementioned 1737 embryo samples were input into various machine learning models, including XGBoost, for training and evaluation. As shown in Table 3, the performance of each model was measured by multiple indicators, including accuracy, F1 score, sensitivity, specificity, positive predictive value, negative predictive value, AUC, and PR-AUC. Among them, the XGBoost model performed best, with an accuracy of 0.641, an F1 score of 0.581, an AUC of 0.780, and a PR-AUC of 0.651, verifying the effectiveness of the multi-source feature system constructed by this method in predicting the genetic testing outcomes of embryos.

[0044] Table 3. Predictive performance evaluation metrics for 9 machine learning models

[0045] Note: AUC: Area under the ROC curve; PR-AUC: Area under the precision-recall curve.

[0046] Furthermore, to explain the model's decision-making logic and identify key predictive factors, the SHAP (SHapley Additive exPlanations) method was used to perform feature importance analysis on the optimal XGBoost model. For example... Figure 2 As shown, the key features of the top 20 contributors are displayed, providing visual support for clinicians to understand the basis of the model's judgment.

[0047] In addition, to verify the model's generalization ability and time stability, the XGBoost model was tested on an independent time validation set. Figure 3 With confusion matrix, Figure 4 The predictive performance of the model on the independent test set was intuitively displayed in two forms: ROC curves, which further confirmed the reliability and robustness of the model in real clinical application scenarios.

[0048] Regarding model interpretability, this invention further demonstrates the model's local decision-making process for a single embryo sample based on the SHAP method. For example... Figures 5 to 7 As shown, the model's performance on a single transplantable embryo is illustrated in waterfall plots. Figure 5 One embryo requiring genetic counseling ( Figure 6 ) and a non-transferable embryo ( Figure 7 The prediction basis is shown in the figure. Each bar represents the direction and magnitude of a feature's contribution to the prediction result, making the black-box decision-making process of the model transparent. This helps clinicians understand the prediction logic of specific cases and improves the trust in human-machine collaboration.

[0049] S3. The outcome prediction model is optimized using an offline reinforcement learning mechanism. This optimization includes: acquiring feedback samples, which are associated with the multi-source features, the three-class prediction probability, the prediction uncertainty, and the target outcome label; updating the action by the policy network based on the output parameters of the current state, where the current state is jointly represented by the latent representation of the multi-source features, the three-class prediction probability, and the prediction uncertainty; training the policy network in an offline playback environment based on a multi-objective reward function, which balances prediction accuracy, probability calibration, and update stability, and controls the model update magnitude through a joint constraint mechanism; and obtaining the updated outcome prediction model parameters based on the trained policy network.

[0050] This step constitutes the core closed-loop optimization mechanism of this invention, which addresses the problem that traditional static prediction models have fixed parameters after deployment and are difficult to continuously optimize using newly added real-world detection results. Its specific implementation process is detailed below.

[0051] First, feedback samples are constructed. After the target embryo completes preimplantation genetic testing, the actual test results, necessary clinical management outcomes, and follow-up information for that embryo are collected and correlated with the inputs and outputs of the prediction phase to form a feedback sample. A single feedback sample is represented as follows: ; in, Indicates the first Multi-source input feature vectors of one embryo; This represents the three-class classification probability output by the prediction model; This indicates the uncertainty of the sample. Indicates the actual test result label; This indicates additional feedback information, including the doctor's correction comments, the time period from which the sample was taken, etc. All historical feedback samples constitute the offline feedback dataset. ; By linking the input features, prediction probabilities, and uncertainty information in the prediction phase with subsequent actual test results, clinical treatment results, and follow-up information, feedback samples and offline feedback datasets are constructed, providing an empirical data foundation for the continuous updating of the model.

[0052] Next, reinforcement learning state modeling is performed. This invention models the model feedback optimization process as a sequential decision problem, encoding the model's current predicted state, sample risk state, and historical optimization state together as the reinforcement learning state. (Time step) The state is defined as: ; Indicated by the feature encoder Extracted low-dimensional latent representation; Outputs the probability for three-class classification; The uncertainty is... The prediction error signal for the current sample can be composed of classification error, cross-entropy loss, or deviation from the true class prediction probability. Thus, when the policy network outputs an action, it can simultaneously perceive the semantic information, prediction confidence, and risk status of the sample.

[0053] Next, the action space is designed. In reinforcement learning, an action is defined as a control action that optimizes the internal parameters of the prediction system; the action space... A For a finite discrete set: ; in: Keep the current model parameters unchanged; : Perform small-step parameter updates; This triggers a conservative rollback strategy, recording only feedback samples without updating the online model. This flexible approach makes the model update process more controllable, avoiding the risks of direct online trial and error.

[0054] Based on this, a multi-objective reward function is constructed. To ensure that the reinforcement learning policy balances prediction accuracy, probability calibration, and update stability, the time step... The reward is defined as: ; in, For accuracy-related benefits; To calibrate the benefit term for probability; To conservatively update the constraint terms; , These are the weighting coefficients.

[0055] The accuracy gain term is defined as: ; This design allows the model to focus not only on whether the classification is correct, but also on the confidence interval between the correct class and the competing classes.

[0056] The probability calibration benefit term, defined to avoid the model prioritizing classification accuracy while neglecting probability reliability, is as follows: ; The conservative update constraint, defined to prevent large single model updates from causing online performance fluctuations, is as follows: ; in, Indicates time Model parameters; Indicates the current optimization strategy; This is the Kullback-Leibler divergence. This term is used to constrain the new strategy from deviating too far from the old strategy, thereby enhancing the controllability of updates in medical scenarios.

[0057] The design of the above reward function, through weighted combination, guides the policy network to automatically balance performance improvement and safety stability during optimization, effectively solving the performance drift and safety risks that are prone to occur in model updates in medical scenarios.

[0058] The learning process of this invention employs a conservative offline reinforcement learning framework. The state-action value function is defined as: ; , where is the discount factor, ranging from 0.8 to 0.99. To improve offline training stability, the following objective function is minimized: ; in, The value network to be trained; The target network is [the network in question]. In the alternative scheme, the reinforcement learning algorithm can be replaced by a conservative offline RL implementation after behavior cloning initialization, as long as offline training and updating are satisfied.

[0059] To avoid performance drift caused by reinforcement learning optimization alone, this invention employs a joint training mechanism combining supervised prediction loss and reinforcement learning value loss. This joint constraint mechanism is achieved by minimizing the total loss function. ; in, The three-class supervision loss is specifically the weighted cross-entropy loss; To reinforce the loss of learning value; This is a balancing factor. By maintaining supervision of the basic classification task while optimizing the reinforcement learning objective, we ensure that the model does not forget its basic classification abilities during the learning of new knowledge, thus maintaining a performance baseline and further enhancing the safety of model updates.

[0060] After offline reinforcement learning training, the system outputs an update strategy. : ; Based on this optimal strategy, the system generates new model parameters. The new parameter combination is only allowed to replace the current model version when it simultaneously meets the conditions of improved returns and controlled update magnitude in an offline playback environment. This safety gating mechanism avoids model oscillations and high-risk erroneous updates that may result from simple incremental training.

[0061] Building upon this foundation, this method also includes version management and audit trail steps. Each model version is assigned an independent version number, and its training time, input data range, feature processing parameters, model parameters, evaluation results, and deployment status are recorded. The system simultaneously records the input information, prediction results, uncertainty, actual detection outcome, feedback time, and version attribution for each embryo sample, forming a traceable audit log. Through this mechanism, the model version origin of any prediction result can be traced, performance changes of different versions over different time periods can be compared, and the input, output, and update process of abnormal results can be traced back, providing verifiable evidence for subsequent model optimization.

[0062] Finally, the system output includes: (1) the three-category prediction results of the target embryo; (2) the prediction probability of each category; (3) uncertainty information; (4) the current model version number and necessary explanatory information, including key features that affect the current prediction results and their degree of contribution, which are used to help doctors understand the basis of the model judgment and improve clinical acceptability.

[0063] Example 2 like Figure 8 As shown, this embodiment provides a preimplantation genetic testing outcome prediction system, including: a feature acquisition module, used to acquire multi-source features of the embryo, wherein the multi-source features include genetic disease features, parental clinical features, ovulation induction cycle features, and embryonic development features; The outcome prediction module is used to input the multi-source features into the outcome prediction model to obtain the three-class prediction probability and prediction uncertainty. The model optimization module is used to perform closed-loop optimization of the outcome prediction model based on an offline reinforcement learning mechanism. The closed-loop optimization includes: acquiring feedback samples, which are associated with the multi-source features, the three-class prediction probabilities, the prediction uncertainty, and the target outcome label; updating the action by the policy network outputting parameters according to the current state, where the current state is jointly represented by the latent representation of the multi-source features, the three-class prediction probabilities, and the prediction uncertainty; training the policy network based on a multi-objective reward function in an offline playback environment, where the multi-objective reward function takes into account prediction accuracy, probability calibration, and update stability, and controls the model update magnitude through a joint constraint mechanism; and obtaining the updated outcome prediction model parameters based on the trained policy network.

[0064] Example 3 Embodiment 3 of the present invention provides an electronic device.

[0065] An electronic device includes a memory, a processor, and a program stored in the memory and running on the processor. The processor includes, but is not limited to, at least one of a central processing unit (CPU), a graphics processing unit (GPU), a neural network processor (NPU), a tensor processor (TPU), or an artificial intelligence acceleration chip. The program is used to execute the steps in the preimplantation genetic testing outcome prediction method as described in Embodiment 1 of the present invention.

[0066] The detailed steps are the same as those for the preimplantation genetic testing outcome prediction method provided in Example 1, and will not be repeated here.

[0067] Example 4 Embodiment 4 of the present invention provides a computer-readable storage medium.

[0068] A computer-readable storage medium having a program stored thereon, which, when executed by a processor, implements the steps in the preimplantation genetic testing outcome prediction method as described in Embodiment 1 of the present invention.

[0069] The detailed steps are the same as those for the preimplantation genetic testing outcome prediction method provided in Example 1, and will not be repeated here.

[0070] Example 5 Embodiment 5 of the present invention provides a computer program product.

[0071] A computer program product includes software code, wherein the program in the software code performs the steps in the preimplantation genetic testing outcome prediction method as described in Embodiment 1 of the present invention.

[0072] The detailed steps are the same as those for the preimplantation genetic testing outcome prediction method provided in Example 1, and will not be repeated here.

[0073] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code. The solutions in the embodiments of the present invention can be implemented using various computer languages. For example, in one implementation, the methods and systems can be developed based on deep learning frameworks (such as TensorFlow, PyTorch, etc.) and using the Python language. Those skilled in the art will understand that other suitable programming languages ​​or tools can also be used for implementation without departing from the core ideas of the present invention.

[0074] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, as well as combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0075] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0076] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0077] The above description is merely a preferred embodiment of this practice and is not intended to limit the scope of this practice. Various modifications and variations can be made to this practice by those skilled in the art. Any modifications, equivalent substitutions, or improvements made within the spirit and principles of this practice should be included within the protection scope of this practice.

Claims

1. A method for predicting the outcome of preimplantation genetic testing, characterized in that, include: Obtain multi-source characteristics of the embryo, including genetic disease characteristics, parental clinical characteristics, ovulation induction cycle characteristics, and embryonic development characteristics; The multi-source features are input into the outcome prediction model to obtain the three-class prediction probability and prediction uncertainty. The outcome prediction model is optimized using an offline reinforcement learning mechanism. The optimization includes: acquiring feedback samples, which are associated with the multi-source features, the three-class prediction probability, the prediction uncertainty, and the target outcome label; updating the action by the policy network based on the output parameters of the current state, where the current state is jointly represented by the latent representation of the multi-source features, the three-class prediction probability, and the prediction uncertainty; training the policy network in an offline playback environment based on a multi-objective reward function, which balances prediction accuracy, probability calibration, and update stability, and controls the model update magnitude through a joint constraint mechanism; and obtaining the updated outcome prediction model parameters based on the trained policy network. The multi-objective reward function is expressed as: ； in, For accuracy-related benefits, To calibrate the benefit term for probability, To conservatively update the constraint terms, , These are the weighting coefficients; The accuracy gain term is defined as follows: ； in, For predicting categories, For the real category, This is the penalty coefficient for misclassification; The probability calibration benefit term is defined as: ； in, For the model to the true category The predicted probability; The conservative update constraint term is defined as follows: ； in, Indicates time Model parameters; Indicates the current optimization strategy; The Kullback-Leibler divergence; For a moment The state.

2. The method according to claim 1, characterized in that, The multi-source features are obtained in the following manner: Collect genetic disease characteristics including inheritance pattern, pathogenic locus type, and paternal or maternal origin information; Clinical characteristics of the parents were collected, including maternal age, anti-Müllerian hormone level, number of antral follicles, body mass index, paternal age, and semen parameters. Collect information on ovulation induction cycle characteristics, including ovulation induction protocol, starting dose, total dose, number of days of ovulation induction, and number of oocytes retrieved; The data collected included embryonic development and grading characteristics such as fertilization method, cleavage stage development status, blastocyst grade, culture days, and embryo morphology scoring. The continuous variables in the genetic disease characteristics, parental clinical characteristics, ovulation induction cycle characteristics, and embryonic development and grading characteristics are standardized, the categorical variables are encoded, and missing values ​​are filled in, and then combined to form the multi-source feature vector.

3. The method according to claim 1, characterized in that, The prediction uncertainty is calculated based on the entropy uncertainty of the three-class prediction results, and the calculation formula is as follows: ； in, This represents the predicted probability for the corresponding category.

4. The method according to claim 1, characterized in that, The joint constraint mechanism is achieved by minimizing the total loss function, which is: ; in, Losses due to three categories of supervision, For balance coefficient, To reinforce the loss of learning value; The reinforcement learning value loss is obtained by minimizing the Bellman residual of the state-action value function, expressed as: ； in, For the value network to be trained, For the target network, This is the discount factor.

5. A preimplantation genetic testing outcome prediction system, characterized in that, include: The feature acquisition module is used to acquire multi-source features of the embryo, including genetic disease features, parental clinical features, ovulation induction cycle features, and embryonic development features. The outcome prediction module is used to input the multi-source features into the outcome prediction model to obtain the three-class prediction probability and prediction uncertainty. The model optimization module is used to perform closed-loop optimization of the outcome prediction model based on an offline reinforcement learning mechanism. The closed-loop optimization includes: acquiring feedback samples, which are associated with the multi-source features, the three-class prediction probability, the prediction uncertainty, and the target outcome label; updating the action by the policy network outputting parameters based on the current state, where the current state is jointly represented by the latent representation of the multi-source features, the three-class prediction probability, and the prediction uncertainty; training the policy network based on a multi-objective reward function in an offline playback environment, where the multi-objective reward function balances prediction accuracy, probability calibration, and update stability, and controls the model update magnitude through a joint constraint mechanism; and obtaining the updated outcome prediction model parameters based on the trained policy network. The multi-objective reward function is expressed as: ； in, For accuracy-related benefits, To calibrate the benefit term for probability, To conservatively update the constraint terms, , These are the weighting coefficients; The accuracy gain term is defined as follows: ； in, For predicting categories, For the real category, This is the penalty coefficient for misclassification; The probability calibration benefit term is defined as: ； in, For the model to the true category The predicted probability; The conservative update constraint term is defined as follows: ； in, Indicates time Model parameters; Indicates the current optimization strategy; The Kullback-Leibler divergence; For a moment The state.

6. An electronic device comprising a memory, a processor, and a computer program stored in the memory and running on the processor, characterized in that, When the processor executes the program, it implements the steps of the method for predicting the outcome of preimplantation genetic testing for embryos according to any one of claims 1 to 4.

7. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the program is executed by the processor, it implements the steps of the method for predicting the outcome of preimplantation genetic testing for embryos according to any one of claims 1 to 4.

8. A computer program product, comprising software code, characterized in that, The program in the software code performs the steps of the method for predicting the outcome of preimplantation genetic testing for embryos according to any one of claims 1 to 4.

Images