A tumor radiotherapy adverse reaction prediction system based on federated learning

Abstract

The application provides a tumor radiotherapy adverse reaction prediction system based on federated learning, relates to the field of artificial intelligence medical treatment, and comprises the following steps: analyzing data quantity ratio, data reliability and data difference of a plurality of training data sets respectively, and evaluating and obtaining a plurality of data contribution degrees; setting a first model parameter aggregation strategy according to the plurality of data contribution degrees, aggregating a plurality of model parameters, constructing a first adverse reaction prediction model and respectively deploying the first adverse reaction prediction model to a plurality of training points; performing iterative cycle training until a preset iteration cycle number is met, and outputting an adverse reaction prediction model for prediction. The application aims to solve the technical problem that the existing prediction system lacks precise measurement of the actual contribution of each participant in the model parameter aggregation process, thereby affecting the prediction accuracy and generalization ability of the model. By introducing a multi-dimensional data contribution degree evaluation mechanism and dynamically setting a model parameter aggregation strategy, the accuracy and generalization ability of the adverse reaction prediction model can be significantly enhanced.

Description

Technical Field

[0001] The present invention relates to the field of artificial intelligence medicine, and in particular to a system for predicting adverse reactions to tumor radiotherapy based on federated learning. Background Art

[0002] Federated learning, a distributed machine learning method that can achieve multi-party collaborative modeling while ensuring data privacy, has been attempted to be applied to medical prediction tasks. This method allows medical institutions to train models based on local data without sharing the original data, and form a unified global model through federated parameter aggregation, thereby improving model performance while protecting patient privacy.

[0003] However, most existing federated learning-based tumor radiotherapy adverse reaction prediction systems adopt fixed or average model parameter aggregation strategies, which fail to fully consider the differences between participating institutions in training data quality, data volume, and data distribution. This one-size-fits-all aggregation method cannot reflect the actual training value of data from all parties during the model fusion process, which can easily lead to limited global model performance, thereby affecting prediction accuracy and the model's generalization ability for different data scenarios, and is difficult to meet the actual needs of tumor radiotherapy clinical practice for high-reliability and personalized predictions. Summary of the Invention

[0004] The purpose of the present invention is to provide a tumor radiotherapy adverse reaction prediction system based on federated learning, so as to solve the technical problem that the existing tumor radiotherapy adverse reaction prediction system based on federated learning fails to fully consider the differences in quality, quantity and distribution of training data among different medical institutions, resulting in a lack of accurate measurement of the actual contribution of each participant in the process of model parameter aggregation, thereby affecting the model prediction accuracy and generalization ability. The tumor radiotherapy adverse reaction prediction system includes a federal coordination server and several training points, and the several training points are respectively deployed in several medical institutions. It also includes: a model parameter output module, which is used to collect several training data sets from several medical institutions in the first training cycle, and supervise the preset model structures of several training points until the preset convergence is reached. Conditions, output several model parameters, and transmit them to the federal coordination server; a data contribution evaluation module, used to perform data volume ratio, data reliability and data difference analysis on the several training data sets at the several training points, evaluate and obtain several data contributions, and transmit them to the federal coordination server; a model parameter aggregation module, used to set a first model parameter aggregation strategy on the federal coordination server according to the several data contributions, generate a first global model parameter after aggregating the several model parameters, construct a first adverse reaction prediction model and decentralize it to the several training points respectively; an adverse reaction prediction module, used to perform iterative cycle training until the preset number of iterative cycles is met, output an adverse reaction prediction model, and perform subsequent tumor radiotherapy adverse reaction prediction.

[0005] Preferably, the tumor radiotherapy adverse reaction prediction system based on federated learning also includes: randomly selecting a first medical institution, collecting patient basic information, patient tumor characteristics and radiotherapy-related characteristics of multiple patients in the first medical institution during the first training cycle, and obtaining a first training input data set, wherein the radiotherapy-related characteristics include radiotherapy equipment type, radiotherapy operating parameters and radiotherapy process; according to a preset adverse reaction type table, collecting adverse reaction data of multiple patients in the first medical institution during the first training cycle, and obtaining a first training supervision data set; obtaining a first training data set based on the mapping combination of the first training input data set and the first training supervision data set, and analyzing in sequence to obtain several training data sets of several medical institutions.

[0006] Preferably, the tumor radiotherapy adverse reaction prediction system based on federated learning also includes: the preset model structure is a composite machine learning model architecture, including a convolutional neural network module and a multi-layer perceptron module, and is jointly modeled through a feature fusion layer, wherein the convolutional neural network module is used to extract image data, and the multi-layer perceptron module is used to extract structured data.

[0007] Preferably, the tumor radiotherapy adverse reaction prediction system based on federated learning also includes: randomly selecting a first training data set of a first training point, and setting the ratio of the data volume of the first training data set to the total data volume of the several training data sets as a first data volume ratio; counting the prediction accuracy of the local training model of the first training point on the validation set, and setting it as a first data reliability coefficient; performing data difference analysis on the first training data set, and outputting a first data generalization coefficient; after dimensionless processing of the first data volume ratio, the first data reliability coefficient and the first data generalization coefficient, weighted evaluation is performed to determine the first data contribution, and added to the several data contribution degrees, wherein the data contribution degree is positively correlated with the data volume ratio, the data reliability coefficient and the data generalization coefficient.

[0008] Preferably, the tumor radiotherapy adverse reaction prediction system based on federated learning also includes: using principal component analysis to reduce the data dimension of the first training data set, and mapping it to a two-dimensional space to draw a two-dimensional scatter plot; based on the data point distribution of the two-dimensional scatter plot, calculating the discreteness index of the point distribution, and outputting the first discreteness as the first data generalization coefficient.

[0009] Preferably, the tumor radiotherapy adverse reaction prediction system based on federated learning also includes: setting the ratio of the data contribution to the sum of several data contributions as the parameter aggregation weight to obtain several parameter aggregation weights; and setting the first model parameter aggregation strategy according to the several parameter aggregation weights.

[0010] Preferably, the tumor radiotherapy adverse reaction prediction system based on federated learning also includes: performing iterative cycle training until a preset number of iterative cycles is met, aggregating several final model parameters, constructing a global adverse reaction prediction model, and performing mean analysis on all training input data within the training cycle, and outputting the global training input data mean; within multiple iterative training cycles of the several training points, retaining the model parameters for local iterative training, constructing several adverse reaction prediction branches, and performing mean analysis on several training input data sets of several training points within the training cycle, respectively, and outputting several local training input data means; constructing an integrated adverse reaction prediction model based on the mapping combination of the several local training input data means and several adverse reaction prediction branches.

[0011] Preferably, the tumor radiotherapy adverse reaction prediction system based on federated learning further comprises: collecting the basic information, tumor characteristics and radiotherapy related characteristics of a to-be-predicted user as to-be-input data; performing data difference analysis on the to-be-input data and the global training input data mean, and outputting a data difference degree; if the data difference degree is less than or equal to a preset difference degree threshold, inputting the to-be-input data into the global adverse reaction prediction model to predict the tumor radiotherapy adverse reaction of the user; and if the data difference degree is greater than the preset difference degree threshold, inputting the to-be-input data into the integrated adverse reaction prediction model to predict the tumor radiotherapy adverse reaction of the user.

[0012] Preferably, the tumor radiotherapy adverse reaction prediction system based on federated learning further comprises: respectively performing similarity comparison on the to-be-input data and the plurality of local training input data means, and outputting a plurality of similarity degrees; in the integrated adverse reaction prediction model, selecting adverse reaction prediction branches corresponding to maximum similarity degrees in a preset proportion, predicting the tumor radiotherapy adverse reaction of the user according to the to-be-input data, and outputting a plurality of preliminary prediction results, wherein the preset proportion is less than 30%; and constructing an adverse reaction prediction data interval according to the plurality of preliminary prediction results as a final prediction result.

[0013] The embodiments of the present application have the following advantages:

[0014] By collecting a plurality of training data sets of a plurality of medical institutions in a first training period, a preset model structure of a plurality of training points is supervised and trained respectively until a preset convergence condition is reached, a plurality of model parameters are output, and are transmitted to a federated coordination server; then, at the plurality of training points, data volume ratio, data reliability and data difference analysis are performed on the plurality of training data sets, a plurality of data contribution degrees are evaluated and obtained, and are transmitted to the federated coordination server; further, at the federated coordination server, a first model parameter aggregation strategy is set according to the plurality of data contribution degrees, the plurality of model parameters are aggregated to generate first global model parameters, a first adverse reaction prediction model is constructed, and is respectively distributed to the plurality of training points; then, iterative period training is performed until a preset number of iteration periods is met, an adverse reaction prediction model is output, and subsequent tumor radiotherapy adverse reaction prediction is performed. That is, by introducing a multi-dimensional data contribution degree evaluation mechanism based on data quality, quantity and distribution difference analysis, a model parameter aggregation strategy is dynamically set, the differential weighting fusion of the training results of each medical institution can be realized, the adaptability of the global model to multi-source heterogeneous training data is effectively improved, thereby the accuracy and generalization ability of the adverse reaction prediction model can be significantly enhanced under the premise of ensuring data privacy, and the actual demand for individualization and high reliability prediction in the tumor radiotherapy clinical scene is met. BRIEF DESCRIPTION OF DRAWINGS

[0015] Figure 1 This is a schematic diagram of the structure of a system for predicting adverse reactions to tumor radiotherapy based on federated learning in the present invention;

[0016] Figure 2 This is a flowchart of the steps of outputting an adverse reaction prediction model in a tumor radiotherapy adverse reaction prediction system based on federated learning of the present invention.

[0017] Description of reference numerals:

[0018] Model parameter output module 11, data contribution evaluation module 12, model parameter aggregation module 13, adverse reaction prediction module 14. DETAILED DESCRIPTION

[0019] The present invention provides a system for predicting adverse reactions to tumor radiotherapy based on federated learning, which solves the technical problem that the existing system for predicting adverse reactions to tumor radiotherapy based on federated learning fails to fully consider the differences in quality, quantity and distribution of training data between different medical institutions, resulting in a lack of accurate measurement of the actual contribution of each participant in the process of model parameter aggregation, thereby affecting the model's prediction accuracy and generalization ability. By introducing a multidimensional data contribution evaluation mechanism based on data quality, quantity and distribution difference analysis and dynamically setting the model parameter aggregation strategy, it is possible to achieve differentiated weighted fusion of the training results of each medical institution, effectively improving the adaptability of the global model to multi-source heterogeneous training data, thereby significantly enhancing the accuracy and generalization ability of the adverse reaction prediction model while ensuring data privacy, and meeting the actual needs for individualized and highly reliable predictions in tumor radiotherapy clinical scenarios.

[0020] Below, the technical solutions of the present invention will be clearly and completely described with reference to the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, rather than all the embodiments of the present invention. It should be understood that the present invention is not limited to the example embodiments described herein. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present invention. It should also be noted that, for the convenience of description, only the parts related to the present invention, rather than all, are shown in the accompanying drawings.

[0021] For examples, please see the attached Figure 1 The present invention provides a system for predicting adverse reactions to tumor radiotherapy based on federated learning. The system includes a federated coordination server and a plurality of training points, wherein the plurality of training points are respectively deployed in a plurality of medical institutions, and further includes:

[0022] The model parameter output module 11 is used to collect several training data sets from several medical institutions in the first training cycle, perform supervised training on the preset model structures of several training points respectively until the preset convergence conditions are reached, output several model parameters, and transmit them to the federal coordination server.

[0023] Furthermore, the model parameter output module 11 is further configured to:

[0024] A first medical institution is randomly selected, and basic patient information, patient tumor characteristics, and radiotherapy-related characteristics of multiple patients in the first medical institution during a first training cycle are collected to obtain a first training input data set, wherein the radiotherapy-related characteristics include radiotherapy equipment type, radiotherapy operating parameters, and radiotherapy process; according to a preset adverse reaction type table, adverse reaction data of multiple patients in the first medical institution during the first training cycle are collected to obtain a first training supervision data set; a first training data set is obtained by mapping and combining the first training input data set and the first training supervision data set, and several training data sets of several medical institutions are analyzed in sequence.

[0025] Specifically, federated learning is a distributed machine learning framework that establishes a collaborative training mechanism across multiple data silos, preventing private data from being centrally processed, thereby enabling joint modeling while protecting data privacy and sovereignty. The present invention provides a federated learning-based tumor radiotherapy adverse reaction prediction system designed to enable collaborative modeling across multiple medical institutions without sharing original patient data, thereby improving the accuracy of adverse reaction predictions during radiotherapy and the generalization of the model. The system comprises a federated coordination server and several training nodes. The federated coordination server is deployed at a central node and is responsible for receiving, aggregating, and distributing global model parameters. The server does not directly access the original data from each institution and is solely responsible for the unified coordination and aggregated optimization of model parameters. The training nodes are distributed across the local data environments of multiple medical institutions (e.g., different hospitals). Each training node possesses local patient radiotherapy data, including pathology information, radiotherapy parameters, and historical adverse reaction records, for local model training. Each training node uses local data for supervised learning based on a pre-defined model structure to train local model parameters and transmits the training results (but not the original data) back to the federated coordination server.

[0026] First, any medical institution is randomly selected from several medical institutions as the first medical institution; then, within the first training cycle (e.g., 15 days as a training cycle), relevant information of patients in each medical institution is collected, including basic patient information, patient tumor characteristics, and radiotherapy-related characteristics of multiple patients. Among them, basic patient information includes basic medical data such as basic demographic information (e.g., age, gender), medical history, and physical indicators of the patient, which provides the model with the patient's overall health background; patient tumor characteristics include tumor type (e.g., lung cancer, breast cancer), stage, size, pathological type, and other tumor-related clinical characteristics. These characteristics have an important impact on the effect of radiotherapy and the risk of adverse reactions; radiotherapy-related characteristics include radiotherapy equipment type, radiotherapy operating parameters, and radiotherapy process. Among them, different equipment may have different technical parameters, such as accelerator model and energy level; radiotherapy operating parameters include radiotherapy dose, irradiation angle, number of irradiations, dose distribution, etc.; radiotherapy process includes process information such as treatment plan, segmentation method, and treatment cycle; this information is then used as the first training input data to construct the first training input data set as the feature input of the model.

[0027] Next, based on a pre-defined adverse reaction table (which details the types and levels of adverse reactions that may occur during tumor radiotherapy, such as skin reactions, mucositis, and fatigue), actual adverse reaction data for each patient during the corresponding training cycle was collected. This included indicators such as the occurrence, intensity, and duration of adverse reactions, such as skin radiation reactions (grades 0 to 4); radiation pneumonitis and radiation enteritis; and the time of onset, duration, and severity. This data was used to obtain the first training supervision dataset, which served as the label data for the model. The patient's training input dataset (i.e., feature data) and the training supervision dataset (i.e., label data) were further combined through mapping to obtain a complete training dataset—paired data consisting of features and corresponding labels—for supervised learning training of the model. Following these steps, the same data collection and construction operations were performed on several medical institutions in sequence, forming multiple training datasets with heterogeneous features and labels.

[0028] Furthermore, the model parameter output module 11 is further configured to:

[0029] The preset model structure is a composite machine learning model architecture, which includes a convolutional neural network module and a multi-layer perceptron module, and is jointly modeled through a feature fusion layer, wherein the convolutional neural network module is used to extract image data, and the multi-layer perceptron module is used to extract structured data.

[0030] Specifically, the preset model structure in the present invention is a composite machine learning model architecture that combines two different types of neural network modules: convolutional neural networks and multi-layer perceptrons. The features extracted by the two are jointly modeled through a feature fusion layer, thereby fully leveraging the advantages of multimodal data and improving prediction performance. Prediction of adverse reactions to tumor radiotherapy involves multiple types of data, including image data and structured data. Image data, such as medical images (CT, MRI, PET, etc.), contains rich spatial and texture information and is suitable for feature extraction using convolutional neural networks. Structured data, such as tabular and numerical attribute data such as patient basic information, tumor clinical characteristics, and radiotherapy parameters, is more suitable for processing using multi-layer perceptrons. To fully exploit the useful information in these two types of data, a composite model structure is designed that complements the advantages of the two. The convolutional neural network module extracts hierarchical features such as edges, textures, and shapes from the image through multiple convolutional and pooling layers to form a high-dimensional deep semantic representation. The multi-layer perceptron module is used to process structured non-image data and consists of multiple fully connected layers. Each node in each layer is fully connected to the nodes in the previous layer, and complex patterns are learned through nonlinear activation functions. The feature fusion layer is used to integrate the deep features extracted by the convolutional neural network module and the multi-layer perceptron module respectively to realize the collaborative learning of multimodal information. This composite model structure can effectively extract and integrate key information of images and structured data through the organic combination of the convolutional neural network module and the multi-layer perceptron module, as well as the collaborative modeling of the feature fusion layer, thereby improving the effect of predicting adverse reactions to tumor radiotherapy.

[0031] Then, using the training input data (patient basic information, patient tumor characteristics and radiotherapy-related characteristics) as input and the training supervision data (adverse reaction data) as supervision, several training data sets are used to supervise the training of the preset model structures of several training points, that is, each training point represents an independent medical institution, and is trained using local training data sets. During the training process, first, the input features are fed into the model, and the model outputs the prediction results; then, the error between the prediction results and the true labels is calculated (such as cross entropy, mean square error); the model parameters are further updated through the back propagation algorithm based on the error; multiple training rounds are looped until the preset convergence conditions are met, such as reaching the maximum number of rounds or the model training loss function drops below a certain threshold, and several model parameters are output, where the model parameters are the neural network weights, bias values, etc. in the local model, and several model parameters are transmitted to the federal coordination server, thereby realizing multi-center joint modeling without leaving the local data.

[0032] The data contribution evaluation module 12 is used to perform data volume ratio, data reliability and data difference analysis on the several training data sets at the several training points, evaluate and obtain several data contributions, and transmit them to the federal coordination server.

[0033] Furthermore, the data contribution evaluation module 12 is further configured to:

[0034] Randomly select a first training data set of a first training point, and set the ratio of the data volume of the first training data set to the total data volume of the several training data sets as a first data volume ratio; and count the prediction accuracy of the local training model of the first training point on the validation set, and set it as a first data reliability coefficient.

[0035] Specifically, in federated learning, the data size and quality of different training points (such as multiple hospitals) may vary greatly. If the data is averaged when the model parameters are aggregated, it is easy to dilute high-quality training points and overestimate low-quality training points, thereby affecting the model's prediction performance and generalization ability. Therefore, it is necessary to calculate the data contribution for each training point and measure the weight it should bear in model training from multiple dimensions (data volume, data quality, etc.). First, any training point is randomly selected from the several training points as the first training point, and the first training data set of the first training point is obtained. Then, the ratio of the data volume of the first training data set to the total data volume of the several training data sets is set as the first data volume ratio. On the other hand, the prediction accuracy of the local training model of the first training point on the validation set is calculated and set as the first data reliability coefficient. The data reliability coefficient reflects the learning effect of the local model at that point, and thus reflects the learnability and quality of the data.

[0036] Perform data difference analysis on the first training data set and output a first data generalization coefficient.

[0037] Furthermore, the present invention further comprises the steps of:

[0038] Principal component analysis is used to perform data dimensionality reduction on the first training data set, and the data is mapped to a two-dimensional space to draw a two-dimensional scatter plot; based on the data point distribution of the two-dimensional scatter plot, a discreteness index of the point distribution is calculated, and a first discreteness is output as a first data generalization coefficient.

[0039] Specifically, principal component analysis (PCA) is a linear dimensionality reduction algorithm used to compress high-dimensional feature vectors (such as a dozen or more patient characteristics) into lower dimensions (usually 2 or 3). The reduced data still maintains the main variability structure of the original data (maximizing information retention). PCA is used to reduce the dimensionality of the first training dataset. First, each feature dimension is standardized to eliminate the impact of dimensional differences on the analysis results. Then, the sample covariance matrix is ​​constructed to characterize the linear correlation between the features. The covariance matrix is ​​then subjected to eigenvalue decomposition to obtain a set of eigenvalues ​​and corresponding eigenvectors. The eigenvalues ​​reflect the variance in the direction of each principal component, that is, the main information content of the data in that direction. Then, the first several eigenvectors with the largest eigenvalues ​​are selected as principal components to form a projection basis. The original high-dimensional data is projected into a low-dimensional space, obtaining a low-dimensional representation that maintains the main structure and distribution characteristics of the original data. This low-dimensional representation not only reduces computational complexity but also provides an effective basis for subsequent data visualization and distribution analysis. Then, after performing PCA dimensionality reduction on the high-dimensional training data, the coordinates of each sample in the directions of the first two principal components (i.e., the position in the two-dimensional space) are obtained, and each sample (high-dimensional feature vector) is projected into the two-dimensional space after dimensionality reduction and drawn as a two-dimensional scatter plot. It can be intuitively observed whether the samples are densely clustered (indicating high data similarity).

[0040] Next, based on the data point distribution of the two-dimensional scatter plot, the dispersion index of the point distribution is calculated, that is, the diversity of the training data set in the feature space is measured, reflecting its generalization support ability for the overall model training. For example, the average distance method is used to calculate the Euclidean distance between all data points and the overall centroid (center point), and the average value is taken. The larger the average distance, the greater the dispersion, which means that the data distribution is wider and the generalization ability is stronger. The first dispersion is output as the first data generalization coefficient. The higher the generalization coefficient, the more comprehensive the features contained in the data set, which helps to improve the generalization ability of the global model.

[0041] By analyzing the spatial distribution of sample points in the two-dimensional scatter plot and calculating its discreteness as a quantitative indicator of generalization ability, we can assist in evaluating the representativeness and importance of the training dataset in multi-source federated learning, thereby improving the adaptability of the global model.

[0042] After dimensionless processing of the first data volume ratio, the first data reliability coefficient and the first data generalization coefficient, a weighted evaluation is performed to determine the first data contribution degree, which is added to the several data contribution degrees, wherein the data contribution degree is positively correlated with the data volume ratio, the data reliability coefficient and the data generalization coefficient.

[0043] Specifically, due to the different dimensions and numerical ranges of different indicators, they need to be standardized (dimensionalized) to achieve a reasonable weighted fusion. Specifically, the first data volume ratio, the first data reliability coefficient, and the first data generalization coefficient are dimensionless. For example, each indicator is normalized to the range of 0 to 1. Next, the three standardized indicators are weighted and summed. The indicator weights are determined based on their impact on the data contribution; a greater impact indicates a greater data contribution. This yields the first data contribution. The data contribution is positively correlated with the data volume ratio, the data reliability coefficient, and the data generalization coefficient. This means that improvements in these indicators directly increase the data contribution, thereby assigning a higher weight to the training point during the model parameter aggregation phase. Using the same method used to calculate the first data contribution, the data contribution of each training data set is calculated sequentially. By standardizing and weighting the quantity, quality, and distribution characteristics of the training data, a data contribution metric is constructed that more accurately reflects the value of each training point to model training, making federated model aggregation more fair and effective, and improving the overall performance of the prediction model.

[0044] The model parameter aggregation module 13 is used to set a first model parameter aggregation strategy on the federal coordination server according to the multiple data contribution degrees, generate a first global model parameter after aggregating the multiple model parameters, construct a first adverse reaction prediction model and delegate it to the multiple training points respectively.

[0045] Furthermore, the model parameter aggregation module 13 is further configured to:

[0046] The ratio of the data contribution degree to the sum of several data contribution degrees is set as the parameter aggregation weight to obtain several parameter aggregation weights; and the first model parameter aggregation strategy is set according to the several parameter aggregation weights.

[0047] Specifically, in federated learning, multiple training points (such as different hospitals) locally train their own model parameters. The server needs to aggregate these parameters into a global model. Traditional methods often use simple averaging, but the data quality, quantity, and representativeness of different training points vary. Direct averaging can lead to reduced model performance. Therefore, to more reasonably reflect the contribution of each training point to the global model, the data contribution of each training point is calculated (combined with indicators such as data volume, reliability, and generalization ability) to guide the weight allocation during aggregation. That is, the ratio of the data contribution to the sum of several data contributions is set as the parameter aggregation weight. Several parameter aggregation weights for several training points are obtained, and a first model parameter aggregation strategy is set based on these several parameter aggregation weights. By dynamically adjusting the weights, the model parameters of hospitals with high training point contributions (large data volume, good data quality, and strong generalization) have a greater impact on the global model, while hospitals with low training point contributions have lower weights. This strategy implements dynamic weight allocation based on data contribution, making the model parameter aggregation process more accurate and fair, and effectively improving the predictive performance and generalization ability of the global model.

[0048] Next, a first model parameter aggregation strategy is adopted to aggregate the model parameters to generate first global model parameters. Specifically, these model parameters are weighted and fused according to the aggregation strategy to calculate a global set of model parameters representing the overall performance of the training results of each training point in the current iteration cycle. The aggregated first global model parameters are then used to construct a unified first adverse reaction prediction model. This model has the ability to integrate data training results from multiple medical institutions and, in theory, can more accurately and universally predict adverse reactions to tumor radiotherapy. The global model parameters are then sent back to all participating training points via the federated coordination server. Each training point uses the new global model as initialization parameters to continue the next round of local training, iterating repeatedly to continuously improve model performance.

[0049] The adverse reaction prediction module 14 is used to perform iterative cycle training until a preset number of iterative cycles is met, output an adverse reaction prediction model, and perform subsequent tumor radiotherapy adverse reaction prediction.

[0050] Further, if Figure 2 As shown, the adverse reaction prediction module 14 is further used to:

[0051] The iteration cycle training is performed until a preset iteration cycle number is met, a global adverse reaction prediction model is constructed after aggregation of a plurality of final model parameters, mean analysis is performed on all training input data in the training cycle, and a global training input data mean is output; in the plurality of iteration training cycles of the plurality of training points, local iteration training is performed on the reserved model parameters, a plurality of adverse reaction prediction branches are constructed, and mean analysis is performed on a plurality of training input data sets of the plurality of training points in the training cycle, and a plurality of local training input data means are output; and an integrated adverse reaction prediction model is constructed according to the plurality of local training input data means and the mapping combination of the plurality of adverse reaction prediction branches.

[0052] Specifically, federated learning adopts a multi-round iteration training manner, wherein each round of iteration includes local training of each training point based on current model parameters, and output of new model parameters; the federated coordination server aggregates these parameters to generate updated global model parameters; and then the global model parameters are distributed back to each training point to start the next round of training; the process is repeatedly executed until a preset iteration cycle number (such as 10 cycles) is reached, ensuring that the model is fully trained and converged; after all iterations are completed, the federated coordination server collects the model parameters output by each training point in the last round of training, and fuses these final model parameters according to a pre-set aggregation strategy (such as weighted average based on data contribution), to obtain global model parameters that comprehensively integrate the final training results of each training point. Finally, the global model parameters after aggregation are used to construct a unified global tumor radiotherapy adverse reaction prediction model, which is based on the training results of multiple institutions and multiple heterogeneous data sources, and is superior to any single training point model in prediction accuracy and generalization ability. Then, the federated coordination server performs statistical analysis on the training input data (including patient basic information, tumor characteristics, radiotherapy related characteristics, etc.) provided by all participating medical institutions in the current complete training cycle, calculates the mean (i.e., the average value of each feature) of these data, and uses the mean to characterize the overall distribution characteristics of the training data, to obtain the global training input data mean.

[0053] On the other hand, during the multiple iterative training cycles of the several training points, the model parameters are retained for local iterative training, that is, each training point (such as a different hospital) continuously uses local data to update and train the model parameters during multiple training iteration cycles. Each training point not only performs the training of the current global model, but also retains the model parameter version obtained during the historical training process, thereby obtaining several adverse reaction prediction branches for several training points. The multi-branch structure helps to improve the robustness and diversity of the model and meet more fine-grained or personalized prediction needs. Next, a mean analysis is performed on several training input data sets of several training points in the training cycle, and several local training input data means of the several training points are output. The several local training input data means and the several adverse reaction prediction branches are then mapped and combined to construct an integrated adverse reaction prediction model. The integrated model can more accurately reflect the heterogeneous data characteristics of each medical institution and improve the prediction effect of adverse reactions to tumor radiotherapy.

[0054] Furthermore, the adverse reaction prediction module 14 is further configured to:

[0055] The basic information, tumor characteristics and radiotherapy-related characteristics of the user to be predicted are collected as the data to be input; data difference analysis is performed on the data to be input and the mean of the global training input data, and the data difference is output; if the data difference is less than or equal to the preset difference threshold, the data to be input is input into the global adverse reaction prediction model to predict the user's tumor radiotherapy adverse reaction.

[0056] Specifically, users to be predicted refer to new patients who have not yet received prediction and need to undergo risk assessment of adverse reactions to tumor radiotherapy. First, basic information of the users to be predicted (such as age, gender, weight, height, medical history, etc.), tumor characteristics (such as tumor type, stage, location, degree of differentiation, molecular typing, etc.) and radiotherapy-related characteristics are collected. Radiotherapy-related characteristics refer to the type of radiotherapy equipment used (linear accelerator, intensity-modulated radiotherapy, etc.), operating parameters (dose, number of fractions, irradiation area) and radiotherapy process (whether combined with chemotherapy, treatment plan, etc.) as the data to be input.

[0057] Next, a data difference analysis is performed on the data to be input and the mean of the global training input data. For example, the data difference is calculated using Euclidean distance. The data difference is used to measure the similarity between the data to be predicted and the training data. The smaller the difference, the closer the input data is to the training data distribution, and the higher the credibility of the model prediction, thus obtaining the data difference. A preset difference threshold is then configured. The preset difference threshold is an empirical tolerance limit set or determined through a validation set to ensure that only users with sufficiently high similarity are directly predicted using the global model to ensure the reliability and validity of the prediction results. If the difference is lower than or equal to the threshold, it indicates that the user's characteristics are highly consistent with the training data, and their data can be directly input into the trained global adverse reaction prediction model for prediction.

[0058] If the data difference is greater than a preset difference threshold, the data to be input is input into the integrated adverse reaction prediction model to predict the user's tumor radiotherapy adverse reaction.

[0059] Furthermore, the present application further comprises the following steps:

[0060] A similarity comparison is performed on the data to be input and the mean values ​​of the several local training input data respectively, and several similarities are output; in the integrated adverse reaction prediction model, an adverse reaction prediction branch corresponding to the maximum similarity of a preset ratio is selected, and the user's tumor radiotherapy adverse reaction is predicted based on the data to be input, and multiple preliminary prediction results are output, wherein the preset ratio is less than 30%; an adverse reaction prediction data interval is constructed based on the multiple preliminary prediction results as the final prediction result.

[0061] Specifically, if the data difference is greater than a preset difference threshold, the feature distribution representing the user differs significantly from the overall training data. Directly using a global model for prediction may reduce prediction accuracy, so a more refined comparison of the user data and the training data features of the local (i.e., single training point) model is required to find a model that is closer to the user. Next, a similarity comparison is performed on the input data to be input and the means of the several local training input data. This involves calculating the similarity between the data of the user to be predicted and each local mean to measure the degree of proximity between the user's features and the historical data of each training point. For example, a similarity comparison is performed using Euclidean distance or cosine similarity. Finally, a similarity score is calculated for each local training point, resulting in several similarities.

[0062] In a federated learning scenario, each medical institution (training site) trains a local model branch using its own data, and all these branches form an integrated adverse reaction prediction model. However, due to the varying individual characteristics of users, their data distribution often resembles the training data of some institutions more closely and diverges from others. Therefore, averaging the predictions of all local models dilutes the effect. Precisely selecting the local models that are most similar to the user can improve prediction accuracy. Next, within the integrated adverse reaction prediction model, the adverse reaction prediction branches corresponding to the highest similarity in a preset percentage are selected. The preset percentage is less than 30%. For example, if there are 30 models, only the top 8 with the highest similarity are selected. This has the advantage of excluding models with significant differences in characteristics from the user, preventing prediction results from being contaminated by low-quality reference values, and avoiding the use of all models, thereby improving prediction efficiency and personalization. Multiple adverse reaction prediction branches are then used to predict adverse reactions to tumor radiotherapy for the user based on the input data, outputting multiple preliminary prediction results. Furthermore, within these multiple preliminary prediction results, the minimum and maximum values ​​are extracted to construct an adverse reaction prediction data interval, which serves as the final prediction result.

[0063] Through the similarity-driven dynamic sub-model selection strategy, the personalized prediction problem in heterogeneous data environments is effectively solved; by limiting the number of models used, the prediction accuracy, efficiency and difference adaptability are taken into account, the system's adaptability to atypical cases is improved, and the practicality of the model in real clinical scenarios is further enhanced.

[0064] In summary, the federated learning-based tumor radiotherapy adverse reaction prediction system provided by the present invention has the following technical effects:

[0065] By collecting several training data sets from several medical institutions in the first training cycle, supervised training is performed on the preset model structures of several training points until the preset convergence conditions are reached, and several model parameters are output and transmitted to the federal coordination server; then, at the several training points, the several training data sets are analyzed for data volume ratio, data reliability and data difference, and several data contributions are evaluated and obtained, and transmitted to the federal coordination server; further, on the federal coordination server, a first model parameter aggregation strategy is set according to the several data contributions, and the several model parameters are aggregated to generate the first global model parameters, and a first adverse reaction prediction model is constructed and decentralized to the several training points respectively; then iterative cycle training is performed until the preset number of iterative cycles is met, and the adverse reaction prediction model is output for subsequent tumor radiotherapy adverse reaction prediction. In other words, by introducing a multidimensional data contribution evaluation mechanism based on data quality, quantity and distribution difference analysis, and dynamically setting the model parameter aggregation strategy, it is possible to achieve differentiated weighted fusion of the training results of various medical institutions, effectively improving the adaptability of the global model to multi-source heterogeneous training data, and thus significantly enhancing the accuracy and generalization ability of the adverse reaction prediction model while ensuring data privacy, meeting the actual needs for individualized and high-reliability predictions in clinical scenarios of tumor radiotherapy.

[0066] The above description of the disclosed embodiments is intended to enable one skilled in the art to implement or use the present invention. Various modifications to these embodiments will be readily apparent to one skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the present invention. Therefore, the present invention is not limited to the embodiments shown herein but is intended to conform to the widest scope consistent with the principles and novel features disclosed herein.

[0067] Obviously, those skilled in the art may make various changes and modifications to the present invention without departing from the spirit and scope of the present invention. Thus, if these modifications and variations of the present invention fall within the scope of the present invention and its equivalents, the present invention is intended to include these modifications and variations.

Claims

1. A system for predicting adverse reactions to tumor radiotherapy based on federated learning, characterized in that: The tumor radiotherapy adverse reaction prediction system includes a federal coordination server and several training points, wherein the several training points are respectively deployed in several medical institutions, and further includes: The model parameter output module is used to collect several training data sets from several medical institutions in the first training cycle, perform supervised training on the preset model structures of several training points respectively until the preset convergence conditions are reached, and output several model parameters, which are transmitted to the federal coordination server; a data contribution evaluation module, configured to perform data volume ratio, data reliability, and data difference analysis on the plurality of training data sets at the plurality of training points, evaluate and obtain a plurality of data contributions, and transmit the data contributions to the federated coordination server; a model parameter aggregation module, configured to set a first model parameter aggregation strategy on the federated coordination server according to the plurality of data contribution degrees, generate first global model parameters after aggregating the plurality of model parameters, construct a first adverse reaction prediction model, and respectively delegate the model to the plurality of training points; The adverse reaction prediction module is used to perform iterative cycle training until the preset number of iterative cycles is met, output an adverse reaction prediction model, and perform subsequent tumor radiotherapy adverse reaction prediction; At the plurality of training points, the plurality of training data sets are respectively analyzed for data volume ratio, data reliability, and data difference, and a plurality of data contribution degrees are evaluated and obtained, including: Randomly selecting a first training data set of a first training point, and setting a ratio of a data volume of the first training data set to a total data volume of the plurality of training data sets as a first data volume ratio; Counting the prediction accuracy of the local training model of the first training point on the validation set, setting it as a first data reliability coefficient; Performing data difference analysis on the first training data set and outputting a first data generalization coefficient; After dimensionless processing of the first data volume ratio, the first data reliability coefficient, and the first data generalization coefficient, a weighted evaluation is performed to determine a first data contribution degree, which is added to the plurality of data contribution degrees, wherein the data contribution degree is positively correlated with the data volume ratio, the data reliability coefficient, and the data generalization coefficient; The step of performing data difference analysis on the first training data set and outputting a first data generalization coefficient includes: Principal component analysis is used to perform data dimensionality reduction on the first training data set, and the data is mapped into a two-dimensional space to draw a two-dimensional scatter plot; Calculating a dispersion index of the point distribution based on the data point distribution of the two-dimensional scatter plot, and outputting a first dispersion as a first data generalization coefficient; Among them, the output adverse reaction prediction model includes: Perform iterative cycle training until the preset number of iterative cycles is met. After aggregating several final model parameters, a global adverse reaction prediction model is constructed, and all training input data within the training cycle are subjected to mean analysis to output the global training input data mean; During multiple iterative training cycles of the multiple training points, retaining model parameters for local iterative training, constructing multiple adverse reaction prediction branches, and performing mean analysis on multiple training input data sets of the multiple training points within the training cycle, and outputting multiple local training input data means; Constructing an integrated adverse reaction prediction model according to the plurality of local training input data means and the plurality of adverse reaction prediction branch mapping combinations; Among them, the prediction of adverse reactions to subsequent tumor radiotherapy includes: Collect basic information, tumor characteristics, and radiotherapy-related characteristics of the user to be predicted as input data; Performing data difference analysis on the data to be input and the mean of the global training input data, and outputting a data difference degree; If the data difference is less than or equal to a preset difference threshold, the data to be input is input into the global adverse reaction prediction model to predict the adverse reaction of the user's tumor radiotherapy; If the data difference is greater than a preset difference threshold, the data to be input is input into the integrated adverse reaction prediction model to predict the user's tumor radiotherapy adverse reaction.

2. A system for predicting adverse reactions to tumor radiotherapy based on federated learning according to claim 1, characterized in that: Collect several training data sets from several medical institutions during the first training cycle, including: Randomly selecting a first medical institution, collecting basic patient information, patient tumor characteristics, and radiotherapy-related characteristics of multiple patients in the first medical institution during a first training cycle to obtain a first training input dataset, wherein the radiotherapy-related characteristics include radiotherapy equipment type, radiotherapy operating parameters, and radiotherapy process; According to a preset adverse reaction type table, adverse reaction data of multiple patients of the first medical institution in a first training cycle are collected to obtain a first training supervision data set; A first training data set is obtained by mapping and combining the first training input data set and the first training supervision data set, and several training data sets of several medical institutions are obtained by sequential analysis.

3. The system for predicting adverse reactions to tumor radiotherapy based on federated learning according to claim 1, characterized in that: The preset model structure is a composite machine learning model architecture, which includes a convolutional neural network module and a multi-layer perceptron module, and is jointly modeled through a feature fusion layer, wherein the convolutional neural network module is used to extract image data, and the multi-layer perceptron module is used to extract structured data.

4. The system for predicting adverse reactions to tumor radiotherapy based on federated learning according to claim 1, characterized in that: Setting a first model parameter aggregation strategy according to the plurality of data contribution degrees includes: The ratio of the data contribution to the sum of several data contributions is set as the parameter aggregation weight to obtain several parameter aggregation weights; A first model parameter aggregation strategy is set according to the plurality of parameter aggregation weights.

5. The system for predicting adverse reactions to tumor radiotherapy based on federated learning according to claim 1, characterized in that: Inputting the data to be input into the integrated adverse reaction prediction model includes: Performing similarity comparisons on the data to be input and the means of the plurality of local training input data, respectively, and outputting a plurality of similarities; In the integrated adverse reaction prediction model, selecting an adverse reaction prediction branch corresponding to the maximum similarity of a preset ratio, predicting the user's adverse reaction to tumor radiotherapy based on the input data, and outputting multiple preliminary prediction results, wherein the preset ratio is less than 30%; An adverse reaction prediction data interval is constructed based on the multiple preliminary prediction results as the final prediction result.

Images