A method for predicting the outcome of assisted reproductive technology procedures in an Internet hospital

Abstract

The present invention relates to the field of assisted reproductive technology, and discloses a method for an Internet hospital to predict the results of assisted reproductive technology procedures, including the following steps: S1, multimodal data collection: based on the Internet hospital platform, patient data related to assisted reproductive technology is obtained from multiple data sources; S2, data preprocessing and feature extraction: preprocessing the multimodal data and extracting features; S3, construction of a dynamic time series prediction model: processing patient status data that changes over time through a long short-term memory network, and constructing a time series model based on a recursive neural network. Through online learning and dynamic feedback mechanisms, the model can be adjusted in real time according to the patient's latest data at each treatment stage. Its dynamic adjustment capability ensures that the prediction model can quickly respond to changes in the patient's physiological and psychological state, and generate personalized treatment recommendations, greatly improving the flexibility and success rate of assisted reproductive technology treatment, and significantly improving the clinical application value compared to traditional static models.

Description

Technical Field

[0001] The present invention relates to the field of assisted reproductive technology, and in particular to a method for an Internet hospital to predict the results of assisted reproductive technology procedures. Background Art

[0002] Assisted reproductive technology is a medical method to help infertile patients achieve fertility, including in vitro fertilization, intracytoplasmic sperm injection, etc. The technology has a wide range of applications, and the success rate depends on the patient's physiological, psychological, genetic and dynamic changes during the treatment process. However, existing technologies still face many challenges in predicting the treatment results of assisted reproductive technology, as follows:

[0003] Static models cannot provide real-time prediction and adjustment based on the patient's current status, lack a dynamic feedback mechanism, and cannot provide personalized advice during the treatment process, which reduces the flexibility and effectiveness of successful treatment.

[0004] Existing technologies lack the ability to quantify prediction uncertainty and provide confidence intervals and uncertainty analysis. Doctors and patients are unable to fully assess risks during the decision-making process, resulting in insufficient reliability and transparency in clinical decision-making.

[0005] The existing assisted reproductive technology outcome prediction models ignore the patient's mental health management, fail to combine psychological factors with physiological data, and lack effective psychological intervention strategies, which affects the improvement of the overall treatment success rate.

[0006] Therefore, those skilled in the art provide a method for an Internet hospital to predict the results of assisted reproductive technology procedures to solve the problems raised in the above-mentioned background technology. Summary of the invention

[0007] In view of the deficiencies in the prior art, the present invention provides a method for an Internet hospital to predict the results of assisted reproductive technology procedures to solve the problems raised in the above-mentioned background technology.

[0008] To achieve the above objectives, the present invention is implemented through the following technical solutions: A method for predicting the results of assisted reproductive technology procedures in an Internet hospital, comprising the following steps: S1, multimodal data collection: based on the Internet hospital platform, obtaining patient data related to assisted reproductive technology from multiple data sources;

[0009] S2, data preprocessing and feature extraction: preprocess multimodal data and extract features;

[0010] S3. Construction of dynamic time series prediction model: The patient status data that changes with time is processed by long short-term memory network, and a time series model based on recurrent neural network is constructed to make real-time predictions of patients at different stages in the assisted reproductive technology cycle. The algorithm formula is:

[0011] ,

[0012] in, represents the prediction result at time point t, is the input data at the current time point, The model parameters are updated by minimizing the following loss function:

[0013] ,

[0014] in, represents the value of the loss function, represents the prediction result at time point t, represents the actual result at time point t, represents the length of the time series, represents the set of parameters of the model, Represents the form of loss function, represents the feature vector;

[0015] S4. Generation of personalized prediction model: Based on cluster analysis, patients are classified into different subgroups, and a personalized prediction model is generated for each subgroup in combination with the Bayesian reasoning algorithm. The Bayesian reasoning model is combined with the following formula for personalized prediction:

[0016] ,

[0017] in, For a given patient characteristic The posterior distribution of the success probability is is the characteristic conditional distribution, is the prior probability of success rate, represents the probability of event x occurring;

[0018] S5. Multimodal data fusion: Different types of data inputs are processed through a multi-input deep learning network. Image data is processed through a convolutional neural network, and time series data is processed through a long short-term memory network. Finally, all inputs are fused through a fully connected layer to generate a unified prediction result.

[0019] Preferably, the data includes clinical data, biomarker data, embryonic time-lapse microscopic imaging data, patient lifestyle data, and real-time health monitoring data;

[0020] The characteristics include the patient's ovarian response, embryonic development potential, and psychological health status.

[0021] Preferably, the multimodal data collection step further comprises obtaining the patient's genetic data by gene sequencing technology, combining candidate gene loci with genetic polymorphisms associated with infertility;

[0022] The gene sequencing technology considers the integration of gene data with clinical data and other multimodal data. In order to capture the nonlinear relationship between different data modes, a deep embedding network is introduced. The network embeds gene information into a high-dimensional vector space and performs fusion prediction with other input data. The specific implementation is as follows:

[0023] Assume the patient's genetic data , Indicates the data value of the i-th gene locus, clinical data , hormone levels and other biomarker data , other environmental data ;

[0024] a. Gene data embedding: First, the gene data is mapped to a high-dimensional feature space through an embedding layer using the matrix Perform linear transformation on each gene locus, and the activation function is , the result is:

[0025] ,in, is a learnable weight matrix, is the bias term, is a nonlinear activation function;

[0026] Multimodal data fusion: Process other multimodal data separately through the fully connected layer to obtain:

[0027] , , ,

[0028] in, , , is the weight matrix for different data modes, , , is the corresponding bias term;

[0029] b. Data fusion and feature interaction: embedding data into features Perform feature interaction and fusion, use the self-attention mechanism to capture the dependencies between different modal data, and generate the final joint feature representation :

[0030] , ,

[0031] in, , , are the query, key, and value vectors of multimodal data, obtained from attention weights through training , and finally generate the fused feature vector ;

[0032] c. Personalized prediction model: fused features Enter the personalized prediction layer and use the Bayesian inference model to generate prediction results: ,

[0033] This formula estimates the success probability through a Bayesian network, is a binary categorical variable indicating whether assisted reproductive technology is successful or not. represents the feature distribution given the prediction success or failure condition, is the prior distribution;

[0034] d. Optimization loss function: L2 regularization loss function is used in the optimization process.

[0035] Preferably, the time series prediction model further includes enhancing the performance of the long short-term memory network model through an attention mechanism, so that the model can automatically capture the most important time period information in the patient data, which is expressed by the following formula:

[0036] ,

[0037] in, is the weight of each time point, is the relevance score, represents the exponential sum of scores at time step k, The energy value Perform exponential operation.

[0038] Preferably, the data preprocessing step includes an outlier detection mechanism, based on standardized residual analysis and K-means clustering algorithm, to automatically identify and process abnormal data points that affect the prediction;

[0039] The standardized residual analysis:

[0040] a. Standardize each data point and map the data of each dimension to the same scale. Calculated by the following formula:

[0041] ,in, is the mean of the feature, is the standard deviation of the feature;

[0042] b. Use standardized residual analysis to evaluate the abnormality of each data point and define the residual of each data point :

[0043] ,

[0044] in, is the predicted value;

[0045] c. Calculate the standardized residual: ,in, is the standard deviation of the residual, if If the value exceeds a certain preset threshold, the point is considered to be an outlier.

[0046] Preferably, the K-means clustering algorithm is used to further process outliers in high-dimensional and nonlinear data. The K-means clustering algorithm is used to detect outliers in the data set. The kernel function of the data point is , which is expressed as:

[0047] ,

[0048] in, is the bandwidth parameter of the kernel function;

[0049] Algorithm steps:

[0050] Determination of initial cluster center: through kernel function Calculate the similarity between data points and initialize k cluster centers using a weighted selection strategy ;

[0051] Assign data points to clusters: For each data point , calculate the kernel function distance to each cluster center:

[0052] ,

[0053] Will Assign to the cluster center with the shortest distance;

[0054] Cluster center update: For each cluster , calculate the new cluster center :

[0055] ,

[0056] in, It is clustering The number of data points in ;

[0057] Outlier detection: After iterative updating, the distance between each data point and its cluster center is calculated. If the distance of a data point exceeds the preset outlier threshold , it is marked as an outlier:

[0058] .

[0059] Preferably, the method further comprises an uncertainty assessment step, comprising sampling the parameters of the Bayesian neural network by a Markov chain Monte Carlo algorithm to estimate the confidence interval of the prediction result;

[0060] The Hamiltonian Monte Carlo algorithm introduces the momentum variable , combined with the energy function, generates differential equations to simulate the parameters The evolutionary trajectory of

[0061] Defining the Joint Distribution :

[0062] ,

[0063] Among them, the Hamiltonian for:

[0064] ,

[0065] in, is the potential energy, expressed as the negative logarithmic posterior distribution:

[0066] ,

[0067] and is the kinetic energy, usually taken as a Gaussian distribution:

[0068] ,

[0069] in, is the mass matrix, usually the identity matrix;

[0070] Hamiltonian equations: ,

[0071] Solve the above equations numerically and update and The value of generates candidate samples in the Markov chain;

[0072] Acceptance rate calculation: According to the acceptance-rejection rule, accept the candidate samples The probability is:

[0073] ,

[0074] By iterative sampling, samples from the posterior distribution are obtained .

[0075] Preferably, the method further comprises an explanatory analysis step, including explaining the influence of local features by the LIME algorithm, so that the doctor can view in detail the contribution of each variable in the prediction result of a single patient;

[0076] The LIME algorithm and the advanced interpretive analysis method of the deep learning model make the machine learning model interpretable through feature perturbation and local linear model fitting mechanism, and provide an explanation for the prediction results of assisted reproductive technology. The steps are as follows:

[0077] LIME explains the black-box model by fitting a local linear model. The behavior nearby, assuming the linear model is: ,

[0078] in, is a perturbed sample of the input feature, The corresponding feature The linear coefficient of

[0079] To generate explanations, LIME takes as input Generate perturbation samples nearby. Specifically, generate N perturbation samples , each sample is obtained by the following formula: ,

[0080] in, is a Gaussian distributed random noise vector, representing the original input of disturbance;

[0081] Linear models focus on LIME assigns weights to perturbed samples in the local neighborhood, using the distance metric function To measure the similarity between the perturbed sample and the original input, the Gaussian kernel:

[0082] ,in, Control the bandwidth of the core;

[0083] Use perturbation samples and corresponding black box model output As the target value, a local linear model is fitted by weighted least squares method , minimize the following weighted loss function:

[0084] , by minimizing this weighted loss function, the parameters of the local linear model are solved ;

[0085] The fitted local linear model The coefficient of The corresponding feature The importance of the feature is explained, and the explanation result is expressed as a ranked list of feature importance, or the contribution of each feature to the prediction result is displayed through visualization;

[0086] LIME introduces sparsity constraints, which are implemented by introducing L1 regularization:

[0087] ,

[0088] in, is the regularization coefficient;

[0089] On high-dimensional datasets, LIME combines feature determination methods to optimize the performance of local linear models and calculates feature determination scores for each sample. , such as using the feature determination method based on information gain:

[0090] ,in, Representation characteristics The information gain between the prediction and the model.

[0091] Preferably, the method further includes a model updating and self-learning mechanism, including weighted processing of feedback data by an automatic differential learning algorithm, giving priority to updating recent patient data;

[0092] The automatic differentiation learning algorithm combines sentiment analysis algorithm and reinforcement learning to optimize the patient's psychological state by dynamically adjusting the intervention strategy, continuously evaluates the patient's psychological state in a data-driven manner, and generates a personalized intervention plan. The algorithm formula steps are as follows:

[0093] Assume that the patient's mental state data , among which, each represents the patient’s mental state on day i, extracted based on the patient’s text feedback. The goal of sentiment analysis is to map the mental state data to a sentiment score , indicating the patient’s emotional state;

[0094] For each , apply sentiment analysis algorithms to generate sentiment scores , the formula is as follows:

[0095] ,

[0096] in, is a sentiment analysis model, are model parameters.

[0097] Preferably, the method further includes a strategy optimization method of reinforcement learning, and the main elements in reinforcement learning include:

[0098] State: The patient's mental state, scored using emotion To express;

[0099] Action: Psychological interventions, assuming there is a set of interventions ;

[0100] Reward: The impact of intervention measures on the patient's psychological state after implementation, reward function Defined as improvement in patients' affect scores after the intervention: ,in, The patient is Day's emotion score;

[0101] Policy: represents the probability distribution of determining the best action in a given state;

[0102] Use deep Q network to optimize psychological intervention strategy. Deep Q network is a reinforcement learning algorithm that combines deep learning and Q learning. The goal is to learn the optimal strategy by maximizing the cumulative reward. The specific formula is as follows:

[0103] Q-value function Indicates that in a given emotional state Next Action The expected rewards you can get, is the parameter of the deep neural network, used to approximate the Q value;

[0104] Update the Q-value function through the Bellman equation:

[0105] ,in, It’s an instant reward. is a discount factor, indicating the importance of future rewards, is the parameter of the target network, indicating the updated Q value;

[0106] The Q-value network is optimized by minimizing the following loss function:

[0107] ,

[0108] Among them, the loss function measures the prediction error of the Q value, and the goal is to minimize the difference between the predicted value and the actual reward;

[0109] Once the reinforcement learning model is trained, the system uses the patient’s emotion score at each time point to , determine the optimal psychological intervention strategy, use a greedy strategy to balance exploration and exploitation, and determine the action :

[0110] ,

[0111] in, is the exploration probability.

[0112] The present invention provides a method for an Internet hospital to predict the results of assisted reproductive technology procedures. It has the following beneficial effects:

[0113] 1. The present invention is based on the Internet hospital. Through online learning and dynamic feedback mechanism, the model can be adjusted in real time according to the patient's latest data at each treatment stage. Its dynamic adjustment capability ensures that the prediction model can quickly respond to changes in the patient's physiological and psychological state and generate personalized treatment recommendations, greatly improving the flexibility and success rate of assisted reproductive technology treatment, and significantly improving the clinical application value compared to traditional static models.

[0114] 2. The present invention is based on the Internet hospital and introduces the Bayesian neural network combined with the Markov chain Monte Carlo sampling method to quantify the uncertainty of the prediction and provide a confidence interval for each prediction result, helping patients and doctors understand the accuracy and uncertainty of the prediction results. It also assists more reliable decision-making by quantifying risks, so that doctors can manage risks more confidently when faced with complex assisted reproductive technology results, provide personalized medical advice, and greatly enhance the transparency and scientific nature of clinical decision-making.

[0115] 3. The present invention is based on the Internet hospital. By combining sentiment analysis algorithms with reinforcement learning, the model can analyze the patient's psychological state and generate personalized psychological intervention strategies based on real-time sentiment data. During the assisted reproductive technology treatment process, the patient's psychological state has an important impact on the results. By dynamically analyzing the psychological state and generating corresponding personalized intervention measures, mental health management is innovatively incorporated into assisted reproductive technology prediction and intervention, thereby improving the overall treatment success rate. BRIEF DESCRIPTION OF THE DRAWINGS

[0116] Figure 1 It is a flow chart of the present invention. DETAILED DESCRIPTION

[0117] In order to make the technical personnel in the technical field understand the scheme of the present invention, the technical scheme in the embodiment of the present invention will be clearly and completely described below in combination with the drawings in the embodiment of the present invention. Obviously, the described embodiment is a partial embodiment of the present invention, not a complete embodiment. Based on the embodiment of the present invention, other embodiments obtained by ordinary technicians in the field without creative work should fall within the scope of protection of the present invention.

[0118] The present invention is described in detail below in conjunction with the accompanying drawings:

[0119] Please refer to the attached Figure 1, an embodiment of the present invention provides a method for an Internet hospital to predict the outcome of an assisted reproductive technology procedure, comprising the following steps: S1, multimodal data collection: based on an Internet hospital platform, obtaining patient data related to assisted reproductive technology from multiple data sources; the data includes clinical data, biomarker data, embryo time-lapse microscopic imaging data, patient lifestyle data, and real-time health monitoring data; the characteristics include the patient's ovarian response, embryo development potential, and mental health status;

[0120] S2, data preprocessing and feature extraction: preprocess multimodal data and extract features;

[0121] S3. Construction of dynamic time series prediction model: By processing the patient status data that changes over time through the long short-term memory network, a time series model based on a recursive neural network is constructed to make real-time predictions of patients at different stages of the assisted reproductive technology cycle, and the following innovative algorithm formula is introduced:

[0122] ,

[0123] in, represents the prediction result at time point t, is the input data at the current time point, The model parameters are updated by minimizing the following loss function:

[0124] ,

[0125] in, represents the value of the loss function, represents the prediction result at time point t, represents the actual result at time point t, represents the length of the time series, represents the set of parameters of the model, Represents the form of loss function, represents the feature vector;

[0126] S4. Generation of personalized prediction model: Based on cluster analysis, patients are classified into different subgroups, and personalized prediction models are generated for each subgroup in combination with the Bayesian reasoning algorithm. The Bayesian reasoning model is combined with the following formula for personalized prediction:

[0127] ,

[0128] in, For a given patient characteristic The posterior distribution of the success probability is is the characteristic conditional distribution, is the prior probability of success rate, represents the probability of event x occurring;

[0129] S5. Multimodal data fusion: Different types of data inputs are processed through multi-input deep learning networks. Image data is processed through convolutional neural networks, and time series data is processed through long short-term memory networks. Finally, all inputs are fused through a fully connected layer to generate a unified prediction result.

[0130] S6. Uncertainty assessment: Bayesian neural network is introduced to quantify the uncertainty of each prediction result, and the result distribution under different scenarios is generated through Monte Carlo simulation. The prediction output includes the success rate estimate and its confidence interval;

[0131] S7. Explanatory analysis: Use the SHAP algorithm to explain the prediction results, show the factors affecting the success rate of assisted reproductive technology, and visualize the contribution of each feature to the results;

[0132] S8. Dynamic feedback mechanism: Based on the real-time data in the assisted reproductive technology program, the prediction model is dynamically adjusted to update the patient's success rate prediction value in real time;

[0133] S9. Patient psychological intervention strategy: Based on the patient's mental health data and combined with the prediction results, generate personalized psychological intervention strategies to help manage the patient's emotions and stress;

[0134] S10. Model update and self-learning mechanism: Feedback the treatment results of each assisted reproductive technology cycle to the model for continuous self-learning to optimize future prediction accuracy, and update the model parameters in combination with online learning methods.

[0135] The benefit of S1 is to improve the prediction accuracy of the model and ensure that the model can fully capture the key variables in the assisted reproductive process. Combined with the patient's real-time health monitoring data, and reflecting the patient's dynamic changes, to ensure the real-time and accuracy of the model input; the benefit of S2 is to improve the quality of the data, reduce the interference of noise, and the feature extraction step further ensures the most important information of the model, reducing the computational complexity and the impact of redundant data; the benefit of S3 is to effectively process data and predict the patient's status at different time points; the benefit of S4 is to significantly improve the accuracy of the prediction and provide patients with more detailed treatment recommendations; the benefit of S5 is to make full use of information from different data sources, enhance the comprehensive prediction ability of the model, effectively improve the accuracy of the model, and make the prediction results more comprehensive and accurate; the benefit of S6 is to more comprehensively reflect the prediction The changes in results under different scenarios provide doctors with more flexible decision-making basis; S7 helps doctors understand the driving factors behind the predictions, enhances clinicians' confidence in the model's prediction results, and uses the information to optimize treatment plans; the benefit of S8 is that it dynamically adjusts the prediction model so that the model can be updated in real time according to the patient's latest condition, thereby maintaining the timeliness and accuracy of the prediction results and further improving the success rate of treatment; the benefit of S9 is that it helps manage the patient's emotions and stress during the treatment process, which is crucial to the success of assisted reproductive technology, can improve the success rate of treatment, and improve the patient's psychological experience; the benefit of S10 is to improve the prediction accuracy of future cycles and enhance the long-term adaptability of the model.

[0136] The multimodal data collection step further includes obtaining the patient's genetic data through gene sequencing technology, combining candidate gene loci with genetic polymorphisms associated with infertility;

[0137] Gene sequencing technology considers the integration of gene data with clinical data and other multimodal data. In order to capture the nonlinear relationship between different data modes, a deep embedding network is introduced. This network embeds gene information into a high-dimensional vector space and performs fusion prediction with other input data. The specific implementation is as follows:

[0138] Assume the patient's genetic data , Indicates the data value of the i-th gene locus, clinical data ,

[0139] Hormone levels and other biomarker data , other environmental data ;

[0140] a. Gene data embedding: First, the gene data is mapped to a high-dimensional feature space through an embedding layer using the matrix Perform linear transformation on each gene locus, and the activation function is , the result is:

[0141] ,in, is a learnable weight matrix, is the bias term, is a nonlinear activation function;

[0142] Multimodal data fusion: Process other multimodal data separately through the fully connected layer to obtain:

[0143] , , ,

[0144] in, , , is the weight matrix for different data modes, , , is the corresponding bias term;

[0145] b. Data fusion and feature interaction: embedding data into features Perform feature interaction and fusion, use the self-attention mechanism to capture the dependencies between different modal data, and generate the final joint feature representation :

[0146] , ,

[0147] in, , , are the query, key, and value vectors of multimodal data, obtained from attention weights through training , and finally generate the fused feature vector ;

[0148] c. Personalized prediction model: fused features Enter the personalized prediction layer and use the Bayesian inference model to generate prediction results: ,

[0149] This formula estimates the success probability through a Bayesian network, is a binary categorical variable indicating whether assisted reproductive technology is successful or not. represents the feature distribution given the prediction success or failure condition, is the prior distribution;

[0150] d. Optimize loss function: To ensure the generalization ability of the model, the L2 regularization loss function is used in the optimization process.

[0151] The overall algorithm collects and preprocesses multimodal data, uses a deep embedding network for high-dimensional space mapping, uses a self-attention mechanism to capture nonlinear relationships between data, and generates personalized prediction results through Bayesian reasoning. The final loss function uses L2 regularization to prevent model overfitting and ensure that the model has high generalization and personalized prediction capabilities.

[0152] Through the effective fusion and deep modeling of multimodal data, accurate personalized prediction results can be generated according to the specific circumstances of each patient, greatly improving the accuracy of predicting the success rate of assisted reproductive technology; combined with Bayesian reasoning, the uncertainty of the prediction results can be quantified, and confidence intervals can be generated to help doctors and patients make more reliable decisions; through L2 regularization, it is ensured that the model will not lose its generalization ability due to overfitting a small number of features, and can maintain a high prediction performance on different patient groups.

[0153] The time series prediction model further includes enhancing the performance of the long short-term memory network model through the attention mechanism, so that the model can automatically capture the most important time period information in the patient data, which is expressed by the following formula:

[0154] ,

[0155] in, is the weight of each time point, is the relevance score, represents the exponential sum of scores at time step k, The energy value Perform exponential operation.

[0156] By introducing the attention mechanism, the LSTM's ability to process long time series is enhanced, so that the model can automatically focus on the key moments in the time series, improve the accuracy of time series prediction, increase the interpretability of the model, and provide doctors with clearer prediction basis. Especially in assisted reproductive technology, the impact of different stages on treatment outcomes varies greatly. The attention mechanism can highlight the differences and help develop more personalized treatment plans.

[0157] Improve the ability to focus on key time periods and automatically identify important time points; improve prediction accuracy and model flexibility so that the model can adapt to data over a long time span; enhance interpretability so that the model can explain the time points that are most important for prediction and assist in clinical decision-making; better handle complex dynamic changes in time series and ensure that features useful for prediction can still be effectively captured in long sequence data.

[0158] The data preprocessing step includes an outlier detection mechanism, which automatically identifies and processes abnormal data points that affect the prediction based on standardized residual analysis and K-means clustering algorithm;

[0159] Standardized residual analysis:

[0160] a. Standardize each data point and map the data of each dimension to the same scale. Calculated by the following formula:

[0161] ,in, is the mean of the feature, is the standard deviation of the feature;

[0162] b. Use standardized residual analysis to assess the abnormality of each data point and define the residual of each data point :

[0163] ,

[0164] in, is the predicted value;

[0165] c. Calculate the standardized residual: ,in, is the standard deviation of the residual, if If it exceeds a certain preset threshold, the point is considered an outlier;

[0166] K-means clustering algorithm is used to further process outliers in high-dimensional and nonlinear data. The outliers in the data set are detected using the K-means clustering algorithm. The kernel function of the data point is , which is expressed as:

[0167] ,

[0168] in, is the bandwidth parameter of the kernel function;

[0169] Algorithm steps:

[0170] Determination of initial cluster center: through kernel function Calculate the similarity between data points and initialize k cluster centers using a weighted selection strategy ;

[0171] Assign data points to clusters: For each data point , calculate the kernel function distance to each cluster center:

[0172] ,

[0173] Will Assign to the cluster center with the shortest distance;

[0174] Cluster center update: For each cluster , calculate the new cluster center :

[0175] ,

[0176] in, It is clustering The number of data points in ;

[0177] Outlier detection: After iterative updating, the distance between each data point and its cluster center is calculated. If the distance of a data point exceeds the preset outlier threshold , it is marked as an outlier:

[0178] .

[0179] The algorithm achieves multi-level outlier detection by integrating standardized residual analysis and K-means clustering algorithm, ensuring that abnormal data can be accurately identified and processed;

[0180] First, standardized residual analysis helps identify points that deviate greatly from model predictions, providing a basis for preliminary anomaly detection. Next, the K-means clustering algorithm further detects anomalies in high-dimensional data, which is especially suitable for nonlinear data. The combination of the two can ensure the accuracy and adaptability of anomaly detection.

[0181] The mean clustering algorithm processes high-dimensional and nonlinear data through kernel functions, which can effectively identify abnormal points that are difficult to detect in complex data, making this method more robust when processing multimodal complex data in assisted reproductive technology;

[0182] During the data preprocessing process, the cluster centers are dynamically updated to ensure that the algorithm can better adapt to the data distribution after each iteration, providing a cleaner and more accurate data set for subsequent prediction models, thereby improving the prediction performance of the overall model.

[0183] The uncertainty assessment step further includes sampling the parameters of the Bayesian neural network through the Markov Chain Monte Carlo algorithm to estimate the confidence interval of the prediction results;

[0184] Hamiltonian Monte Carlo algorithm introduces momentum variable , combined with the energy function, generates differential equations to simulate the parameters The evolutionary trajectory of

[0185] Defining the Joint Distribution :

[0186] ,

[0187] Among them, the Hamiltonian for:

[0188] ,

[0189] in, is the potential energy, expressed as the negative logarithmic posterior distribution:

[0190] ,

[0191] and is the kinetic energy, usually taken as a Gaussian distribution:

[0192] ,

[0193] in, is the mass matrix, usually the identity matrix;

[0194] Hamiltonian equations: ,

[0195] Solve the above equations numerically and update and The value of generates candidate samples in the Markov chain;

[0196] Acceptance rate calculation: According to the acceptance-rejection rule, accept the candidate samples The probability is:

[0197] ,

[0198] By iterative sampling, samples from the posterior distribution are obtained .

[0199] The Hamiltonian Monte Carlo algorithm combines the potential energy and kinetic energy in the physical system and uses momentum to simulate the sampling trajectory of the Bayesian neural network parameters. It can generate a more effective Markov chain. By introducing the Hamiltonian equation and the acceptance-rejection rule, the algorithm is more efficient in exploring the parameter space and ensures that the final sample conforms to the posterior distribution.

[0200] Through the Hamiltonian equation and energy model, the algorithm can generate a more efficient sampling path, avoid the random walk problem in Markov chain Monte Carlo, and quickly converge to the target distribution. The mechanism makes the sampling of Bayesian neural networks in high-dimensional parameter space more flexible and accurate;

[0201] Through Hamiltonian Monte Carlo algorithm sampling, the model can generate parameter samples from the Bayesian posterior distribution. The samples are used to estimate the confidence interval of the prediction results. Compared with traditional deterministic neural networks, this method can quantify the uncertainty of the prediction results and provide the confidence interval of the prediction results, making the prediction more credible.

[0202] The Hamiltonian Monte Carlo algorithm's acceptance rate calculation formula can prevent the model from falling into a local optimum. By dynamically adjusting the sampling path, the Hamiltonian Monte Carlo algorithm can explore more parameter spaces and ensure a more comprehensive distribution of parameter samples.

[0203] For Bayesian neural networks with high-dimensional parameters, traditional sampling methods are difficult to handle effectively. The Hamiltonian Monte Carlo algorithm can generate more effective sampling trajectories in the parameter space through momentum guidance, which is especially suitable for complex Bayesian neural network models.

[0204] The explanatory analysis step includes explaining the impact of local features through the LIME algorithm, allowing doctors to view in detail the contribution of each variable in the prediction results of a single patient;

[0205] The advanced interpretable analysis method of the LIME algorithm and deep learning model makes the machine learning model interpretable through feature perturbation and local linear model fitting mechanism, and provides an explanation for the prediction results of assisted reproductive technology. The steps are as follows:

[0206] LIME explains the black-box model by fitting a local linear model. The behavior nearby, assuming the linear model is: ,

[0207] in, is a perturbed sample of the input feature, The corresponding feature The linear coefficient of

[0208] To generate explanations, LIME takes as input Generate perturbation samples nearby. Specifically, generate N perturbation samples , each sample is obtained by the following formula: ,

[0209] in, is a Gaussian distributed random noise vector representing the original input of disturbance;

[0210] Linear models focus on LIME assigns weights to perturbed samples in the local neighborhood, using the distance metric function To measure the similarity between the perturbed sample and the original input, the Gaussian kernel:

[0211] ,in, Control the bandwidth of the core;

[0212] Use perturbation samples and corresponding black box model output As the target value, a local linear model is fitted by weighted least squares method , minimize the following weighted loss function:

[0213] ,

[0214] By minimizing this weighted loss function, the parameters of the local linear model are solved ;

[0215] The fitted local linear model The coefficient of The corresponding feature The importance of the feature is explained, and the explanation result is expressed as a ranked list of feature importance, or the contribution of each feature to the prediction result is displayed through visualization;

[0216] LIME introduces sparsity constraints, which are implemented by introducing L1 regularization:

[0217] ,

[0218] in, is the regularization coefficient;

[0219] On high-dimensional datasets, LIME combines feature determination methods to optimize the performance of local linear models and calculates feature determination scores for each sample. , such as using the feature determination method based on information gain:

[0220] ,

[0221] in, Representation characteristics The information gain between the prediction and the model.

[0222] The LIME algorithm generates interpretable analysis for black-box models such as deep learning through feature perturbation and local linear fitting mechanisms, allowing doctors to understand the predictive behavior of complex models in assisted reproductive technology and quantify the contribution of each feature to the prediction results;

[0223] By generating perturbation samples, LIME fits a local linear model and optimizes the model through weighted least squares, making complex models such as deep learning interpretable in local areas. The linear model coefficient β reflects the importance of each feature to the prediction result.

[0224] By introducing L1 regularization, LIME strengthens the sparsity of local linear models and retains a few of the most important features for interpretation, so that doctors can quickly understand which variables play the most critical role in the prediction of a specific patient;

[0225] By generating perturbation samples near the original input, LIME can capture the subtle behavior of the model in the local area, and ensure that the linear model focuses on the local area through distance weighting, thus enhancing the locality and reliability of the explanation;

[0226] LIME provides a reasonable feature selection mechanism for high-dimensional data by calculating feature information gain, which can further optimize the interpretation performance and ensure that the interpretation results are accurate and have strong practical significance.

[0227] The model update and self-learning mechanism includes weighted processing of feedback data through an automatic differentiated learning algorithm, giving priority to updating recent patient data;

[0228] The automatic differentiation learning algorithm combines sentiment analysis algorithm and reinforcement learning to optimize the patient's psychological state by dynamically adjusting the intervention strategy. It continuously evaluates the patient's psychological state in a data-driven manner and generates personalized intervention plans. The algorithm formula steps are as follows:

[0229] Assume that the patient's mental state data , among which, each represents the patient’s mental state on day i, extracted based on the patient’s text feedback. The goal of sentiment analysis is to map the mental state data to a sentiment score , indicating the patient’s emotional state;

[0230] For each , apply sentiment analysis algorithms to generate sentiment scores , the formula is as follows:

[0231] ,

[0232] in, is a sentiment analysis model, are model parameters;

[0233] The strategy optimization method of reinforcement learning. The main elements of reinforcement learning include:

[0234] State: The patient's mental state, scored using emotion To express;

[0235] Action: Psychological interventions, assuming there is a set of interventions ;

[0236] Reward: The impact of intervention measures on the patient's psychological state after implementation, reward function Defined as improvement in patients' affect scores after the intervention: ,in, The patient is Day's emotion score;

[0237] Policy: represents the probability distribution of determining the best action in a given state;

[0238] Use deep Q network to optimize psychological intervention strategy. Deep Q network is a reinforcement learning algorithm that combines deep learning and Q learning. The goal is to learn the optimal strategy by maximizing the cumulative reward. The specific formula is as follows:

[0239] Q-value function Indicates that in a given emotional state Next Action The expected rewards you can get, is the parameter of the deep neural network, used to approximate the Q value;

[0240] Update the Q-value function through the Bellman equation:

[0241] ,

[0242] in, It’s an instant reward. is a discount factor, indicating the importance of future rewards, is the parameter of the target network, indicating the updated Q value;

[0243] The Q-value network is optimized by minimizing the following loss function:

[0244] ,

[0245] Among them, the loss function measures the prediction error of the Q value, and the goal is to minimize the difference between the predicted value and the actual reward;

[0246] Once the reinforcement learning model is trained, the system uses the patient’s emotion score at each time point. , determine the optimal psychological intervention strategy, use a greedy strategy to balance exploration and exploitation, and determine the action :

[0247] ,

[0248] in, is the exploration probability.

[0249] The algorithm combines sentiment analysis and reinforcement learning to optimize the patient's psychological state by dynamically adjusting the intervention strategy. The sentiment analysis model first converts the patient's psychological state into a sentiment score, which is used as the state input in reinforcement learning. Reinforcement learning then dynamically adjusts the intervention strategy based on the sentiment score and reward function, ultimately generating a personalized psychological intervention plan.

[0250] The sentiment analysis model provides quantified sentiment state input for reinforcement learning, helping the model to dynamically evaluate the patient's psychological state, so that reinforcement learning can generate personalized psychological intervention plans based on different sentiment scores;

[0251] By defining the improvement of the patient's emotional state as a reward, the reinforcement learning model can evaluate the effect of intervention measures in real time. The optimization of the reward function enables the model to gradually identify the intervention measures that are most effective in improving the patient's emotional state.

[0252] Through the deep Q network, the model can estimate the long-term cumulative rewards and optimize the intervention strategy through the Bellman equation. The Q-value function combines the current rewards and the expected future rewards, so that the model can focus on short-term effects and optimize long-term mental health management.

[0253] The greedy strategy helps the model gradually discover the optimal intervention strategy by balancing exploration and utilization. Through the mechanism, the model can explore new strategies to ensure that patients can receive adaptive intervention at different stages.

[0254] Although embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions and variations may be made to the embodiments without departing from the principles and spirit of the present invention, and that the scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. A method for predicting the results of assisted reproductive technology procedures in an Internet hospital, characterized in that: The steps include: S1. Multimodal data collection: Based on the Internet hospital platform, patient data related to assisted reproductive technology are obtained from multiple data sources; S2, data preprocessing and feature extraction: preprocess multimodal data and extract features; S3. Construction of dynamic time series prediction model: The patient status data that changes with time is processed by long short-term memory network, and a time series model based on recurrent neural network is constructed to make real-time predictions of patients at different stages of the assisted reproductive technology cycle. The formula is: in, represents the prediction result at time point t, x(t) is the input data at the current time point, θ is the model parameter, and the model parameter is updated by minimizing the following loss function: Among them, L represents the value of the loss function, represents the prediction result at the tth time point, y(t) represents the actual result at the tth time point, T represents the length of the time series, θ represents the parameter set of the model, L(θ) represents the form of the loss function, and λ represents the feature vector; S4. Generation of personalized prediction model: Based on cluster analysis, patients are classified into different subgroups, and a personalized prediction model is generated for each subgroup in combination with the Bayesian inference algorithm. The Bayesian inference model is combined with the following formula for personalized prediction: Where P(y|x) is the posterior distribution of the probability of success given patient feature x, P(x|y) is the conditional distribution of the feature, P(y) is the prior probability of the success rate, and P(x) represents the probability of event x occurring; S5. Multimodal data fusion: Different types of data inputs are processed through multi-input deep learning networks. Image data is processed through convolutional neural networks, and time series data is processed through long short-term memory networks. Finally, all inputs are fused through a fully connected layer to generate a unified prediction result. The multimodal data collection step further includes: obtaining the patient's genetic data through gene sequencing technology, combining candidate gene loci with genetic polymorphisms associated with infertility; The gene sequencing technology considers the integration of gene data with clinical data and multimodal data. In order to capture the nonlinear relationship between different data modes, a deep embedding network is introduced. The network embeds gene information into a high-dimensional vector space and performs fusion prediction with the input data. The specific implementation is as follows: Assume that the patient's genetic data G = {g1, g2, ..., g n }, g i represents the data value of the i-th gene locus, clinical data C = {c1, c2, ..., c m }, biomarker data B = {b1, b2, ..., b k }, environmental data E = {e1, e2, ..., e p }; a. Gene data embedding: First, the gene data is mapped to a high-dimensional feature space through an embedding layer using the matrix W G Performing linear transformation on each gene locus, with activation function σ(·), the result is: G′=σ(W G ·G+b G ), where W G is the learnable weight matrix, b G is the bias term, σ(·) is the nonlinear activation function; Multimodal data fusion: Multimodal data are processed separately through the fully connected layer to obtain: C′=σ(W C ·C+b C ), B′=σ(W B ·B+b B ), E′=σ(W E ·E+b E ), Among them, W C , W B , W E is the weight matrix of different data modes, b C 、b B 、b E is the corresponding bias term; b. Data fusion and feature interaction: The features G′, C′, B′, and E′ after data embedding are interacted and fused, and the self-attention mechanism is used to capture the dependencies between different modal data to generate the final joint feature representation Z: Among them, q i , k i 、v i are the query, key, and value vectors of multimodal data, obtained from the attention weights α through training i , and finally generate the fused feature vector Z; c. Personalized prediction model: The fused feature Z is input into the personalized prediction layer, and the Bayesian inference model is used to generate the prediction result: This formula estimates the probability of success through a Bayesian network, y is a binary categorical variable for the success or failure of assisted reproductive technology, P(Z|y) represents the characteristic distribution under the given prediction of success or failure, and P(y) is the prior distribution; d. Optimize loss function: L2 regularized loss function is used in the optimization process; The method also includes an uncertainty assessment step, including sampling the parameters of the Bayesian neural network through a Markov chain Monte Carlo algorithm to estimate the confidence interval of the prediction result; The Hamiltonian Monte Carlo algorithm simulates the evolution trajectory of the parameter θ by introducing the momentum variable p and combining it with the energy function to generate a differential equation; Define the joint distribution P(θ, p): P(θ,p)∝exp(-H(θ,p)), Among them, the Hamiltonian H(θ, p) is: H(θ, p) = U(θ) + K(p), Among them, U(θ) is the potential energy, expressed as the negative logarithmic posterior distribution: U(θ)=-logP(D|θ)-logP(θ), K(p) is the kinetic energy, which is usually taken as Gaussian distribution: Where M is the mass matrix, usually the unit matrix; Hamiltonian equations: Solve the above equations by numerical methods, update the values ​​of θ and p, and generate candidate samples in the Markov chain; Acceptance rate calculation: According to the acceptance-rejection rule, the probability of accepting a candidate sample (θ′, p′) is: By iterative sampling, we obtain samples from the posterior distribution {θ (1) ,θ (2) ,…,θ (T) }; The method also includes an explanatory analysis step, including explaining the influence of local features through the LIME algorithm, so that the doctor can view in detail the contribution of each variable in the prediction result of a single patient; The LIME algorithm and the advanced interpretive analysis method of the deep learning model make the machine learning model interpretable through feature perturbation and local linear model fitting mechanism, and provide an explanation for the prediction results of assisted reproductive technology. The steps are as follows: LIME explains the behavior of the black box model near x0 by fitting a local linear model. Assume that the linear model is: Among them, z is the perturbation sample of the input feature, β i is the corresponding feature z i The linear coefficient of To generate explanations, LIME generates perturbation samples around the input x0. Specifically, it generates N perturbation samples {z1, z2, …, z N }, each sample is obtained by the following formula: i =x0+∈ i , in, is a Gaussian distributed random noise vector, representing the perturbation of the original input x0; The linear model focuses on the local neighborhood of x0, while LIME assigns weights to the perturbed samples and uses the distance metric function π(x0, z i ) to measure the similarity between the perturbation sample and the original input, Gaussian kernel: Among them, σ controls the bandwidth of the core; Use the perturbation sample and the corresponding black box model output f(z i ) as the target value, and fit the local linear model g(z) by weighted least squares method to minimize the following weighted loss function: By minimizing this weighted loss function, the parameters β of the local linear model are solved. i ; The coefficient β of the fitted local linear model g(z) i That is, it is the importance explanation of the corresponding feature x0. The explanation result is expressed as a ranking list of feature importance, or the contribution of each feature to the prediction result is displayed through visualization; LIME introduces sparsity constraints, which are implemented by introducing L1 regularization: Where λ is the regularization coefficient; On high-dimensional data sets, LIME combines feature determination methods to optimize the performance of local linear models and calculates the feature determination score S for each sample. i , such as using the feature determination method based on information gain: Among them, I(·,·) represents the feature Information gain with the model prediction; The method also includes a model updating and self-learning mechanism, including weighting the feedback data through an automatic differential learning algorithm to give priority to updating recent patient data; The automatic differentiation learning algorithm combines sentiment analysis algorithm and reinforcement learning to optimize the patient's psychological state by dynamically adjusting the intervention strategy, continuously evaluates the patient's psychological state in a data-driven manner, and generates a personalized intervention plan. The algorithm formula steps are as follows: Assume that the patient's mental state data S = {s1, s2, ..., s n }, where each s i represents the patient’s psychological state on day i, extracted based on the patient’s text feedback. The goal of sentiment analysis is to map the psychological state data to the sentiment score e i ∈R, represents the patient’s emotional state; For each i , apply sentiment analysis algorithm to generate sentiment score e i , the formula is as follows: e i =f emotion (s i ;θ emotion ), Among them, f emotion is the sentiment analysis model, θ emotion are model parameters; The method also includes a strategy optimization method for reinforcement learning, and the main elements in reinforcement learning include: State: The patient's mental state, scored using emotion i To express; Action: Psychological intervention measures. Assume that there is a set of intervention measures A = {a1, a2, ..., a m }; Reward: The impact of intervention measures on the patient's psychological state, reward function r i Defined as the improvement in the patient's affect score after the intervention: i =e i+1 -e i , where e i+1 is the patient's emotion score on day i+1; Policy: represents the probability distribution of determining the best action in a given state; Use deep Q network to optimize psychological intervention strategy. Deep Q network is a reinforcement learning algorithm that combines deep learning and Q learning. The goal is to learn the optimal strategy by maximizing the cumulative reward. The specific formula is as follows: Q value function Q(e i , a; θ) represents the given emotional state e i The expected reward for taking action a is θ, which is the parameter of the deep neural network and is used to approximate the Q value. Update the Q-value function through the Bellman equation: Among them, r i is the immediate reward, γ∈[0,1] is the discount factor, indicating the importance of future rewards, and γ∈[0,1] is the parameter of the target network, indicating the updated Q value; The Q-value network is optimized by minimizing the following loss function: Among them, the loss function measures the prediction error of the Q value, and the goal is to minimize the difference between the predicted value and the actual reward; Once the reinforcement learning model is trained, the system uses the patient’s emotion score e at each time point. i , determine the optimal psychological intervention strategy, use the greedy strategy to balance exploration and exploitation, and determine action a i : where ∈ is the exploration probability.

2. The method for predicting the results of assisted reproductive technology procedures in an Internet hospital according to claim 1, characterized in that: The data include clinical data, biomarker data, embryo time-lapse microscopy data, patient lifestyle data, and real-time health monitoring data; the characteristics include the patient's ovarian response, embryo development potential, and mental health status.

3. The method for predicting the results of assisted reproductive technology procedures in an Internet hospital according to claim 1, characterized in that: The time series prediction model includes enhancing the performance of the long short-term memory network model through the attention mechanism, so that the model can automatically capture the most important time period information in the patient data, which is expressed by the following formula: Among them, α t is the weight of each time point, e t is the relevance score, represents the exponential sum of the scores at time step k, exp(e t ) is the energy value e t Perform exponential operation.

4. The method for predicting the results of assisted reproductive technology procedures in an Internet hospital according to claim 1, characterized in that: The data preprocessing step includes an outlier detection mechanism, based on standardized residual analysis and K-means clustering algorithm, to automatically identify and process abnormal data points that affect the prediction; The standardized residual analysis: a. Standardize each data point and map the data of each dimension to the same scale. The standardized data X′={x′1,x′2,…,x′ n } is calculated by the following formula: Among them, μ is the mean of the feature, σ is the standard deviation of the feature; b. Use standardized residual analysis to evaluate the abnormality of each data point and define the residual r of each data point i : in, is the predicted value; c. Calculate the standardized residual: Among them, σ r is the standard deviation of the residual, if z i If the value exceeds a certain preset threshold, the point is considered to be an outlier.

5. The method for predicting the results of assisted reproductive technology procedures in an Internet hospital according to claim 4, characterized in that: The K-means clustering algorithm is used to further process outliers in high-dimensional and nonlinear data. The K-means clustering algorithm is used to detect outliers in the data set. The kernel function of the data point is K(x i ,x j ), which is expressed as: Among them, σ is the bandwidth parameter of the kernel function; Algorithm steps: Determination of the initial cluster center: Through the kernel function K(x i ,x j ) calculates the similarity between data points and uses a weighted selection strategy to initialize k cluster centers {μ1, μ2, …, μ k }; Assign data points to clusters: For each data point x i , calculate the kernel function distance to each cluster center: d(x i ,μ j )=K(x i ,x i )-2K(x i ,μ j )+K(μ j ,μ j ), x i Assign to the cluster center with the shortest distance; Cluster center update: For each cluster C j , calculate the new cluster center μ j : Among them, |C j | is cluster C j The number of data points in ; Outlier detection: After iterative updating, the distance between each data point and its cluster center is calculated. If the distance of a data point exceeds the preset anomaly threshold δ, it is marked as an outlier:

Images