Multi-center data processing method based on self-paced learning and personalized federated learning
Through self-step learning and personalized federated learning methods, dynamically adjust sample selection and model parameter separation, the problem of limited model generalization ability and privacy protection caused by medical data heterogeneity is solved, and efficient in a multi-center environment Data processing and model customization.
Patent Information
- Application Number
- CN202411166576.5
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2024-08-23
- Publication Date
- 2024-11-15
- Estimated Expiration
- 2044-08-23
AI Technical Summary
The sharing and utilization of medical data among multiple medical centers is limited by privacy protection, resulting in data silos, and the existing personalized federated learning methods do not perform well in heterogeneous data environments, making it difficult to converge to the global optimal model.
Adopting a global federated learning strategy based on self-step learning, dynamically adjusting the sample selection threshold, training the global federated model, and splitting it into shared feature extractor and private decision-maker parameters, and personalized training is carried out to adapt to the data characteristics of different medical centers. .
It effectively alleviates the negative impact of data heterogeneity on model training, enhances the generalization ability and convergence of the model, and protects patient privacy information, and customizes personalized models for each medical center.
Smart Images

Figure CN118690203B_ABST
Abstract
Description
Technical Field
[0001] The present invention belongs to the technical field of medical data processing, and in particular to a multi-center data processing method based on self-paced learning and personalized federated learning. Background Art
[0002] With the rapid development of medical testing technology and information technology, the means of disease inspection and analysis have undergone tremendous changes, bringing new opportunities for artificial intelligence to assist doctors in decision-making and optimize treatment plans. In the field of intensive care, continuous monitoring of patients generates a large amount of vital signs data, which can provide strong data support for machine learning.
[0003] Traditional machine learning research is usually conducted in a large database. However, because medical data contains a large amount of patient information, in order to protect patient privacy, these data are not allowed to be transferred and shared between multiple medical centers. This results in medical institutions having to keep data locally, thus forming data islands, making it difficult for a large amount of data to be effectively used.
[0004] Federated learning is a distributed machine learning method that can collaboratively train models between multiple clients (medical centers) without sharing sensitive patient data, so it is widely used in the medical field. However, due to data heterogeneity, a unified global federated model cannot meet the personalized requirements of clients and performs poorly on specific clients.
[0005] Personalized federated learning can combine the client's local data for personalized training on the premise of training the global federated model, thereby solving the impact of data heterogeneity on federated learning. Mainstream personalized federated learning can be divided into the following two categories: 1) Global federated model personalization: Establish a powerful global federated model through federated learning, and use local data on the client to personalize the model; 2) Direct learning personalized model: Directly train a personalized model for the client by modifying the model aggregation method during the federated learning process.
[0006] However, the effectiveness of the global federated model personalization method depends largely on the performance of the global federated model. If the global federated model performs poorly in different data environments, it is difficult to significantly improve the model effect even if it is adjusted through local data. In addition, medical data in multi-center scenarios usually exhibit the characteristics of non-independent and identical distribution. Federated aggregation methods represented by the federated averaging algorithm FedAvg are difficult to converge to the global optimal model. Direct learning personalized model methods rely heavily on the degree of statistical heterogeneity of the data. This means that these methods usually only perform well in specific types of heterogeneity. For example, FedPer performs well in cases of heterogeneous label distribution, while LG-FedAvg works better in cases of heterogeneous feature distribution. In real scenarios, due to privacy issues, the data heterogeneity of multiple medical institutions is often difficult to evaluate, which makes it challenging to choose a suitable direct learning personalized federated learning method. Summary of the invention
[0007] To solve the above technical problems, the present invention provides a multi-center data processing method based on self-paced learning and personalized federated learning. By designing a personalized prediction model and performing model training based on a self-paced learning strategy, the negative impact of heterogeneous data on model training is effectively alleviated.
[0008] To achieve the above purpose, the technical solution adopted by the present invention is as follows:
[0009] A multi-center data processing method based on self-paced learning and personalized federated learning, the method comprising the following steps:
[0010] Step 1: Train the global federated model based on the global self-paced learning strategy: Sort and select client training samples based on the training sample loss value and the sample selection threshold, and dynamically adjust the sample selection threshold so that the global federated model selects different client training samples at different training stages;
[0011] Step 2: Build a personalized model for each center: Split the trained global federated model into a feature extractor part and a decision maker part, decouple the parameters of the feature extractor part and the decision maker part, set the decoupled feature extractor parameters as shared parameters, and set the decision maker parameters as client private parameters.
[0012] Furthermore, the step 1 comprises:
[0013] Step 1.1: The server establishes a connection with the client, determines the global federated learning participants, and initializes the global federated model;
[0014] Step 1.2, the server sends the initialized global federated model to the client;
[0015] Step 1.3: The client accepts the initialized global federated model and calculates the single sample loss value for the local training sample data. ;
[0016] Step 1.4: Based on the current number of training rounds t, the total number of training rounds T, and the client local data set The maximum loss value of all training samples And the average loss value Calculate sample selection threshold ;
[0017] ,
[0018] in, Indicates the client local data set when the current training round number is t The average loss value of all training samples, Indicates the client local data set when the current training round number is t The maximum loss value of all training samples;
[0019] Step 1.5: Select a single sample loss value Not greater than the sample selection threshold The local training sample set constitutes a new dataset Training; Step 1.6, based on the new data set The trained client local model is uploaded to the server, and the server aggregates all local models to obtain a new global federated model;
[0020] Step 1.7: Repeat steps 1.2 to 1.6 until the set total number of training rounds T is met to obtain the final global federated model.
[0021] Furthermore, the step 2 comprises:
[0022] Step 2.1: The client accepts the final global federation model;
[0023] Step 2.2, split the final global federated model into a feature extractor including a double-layer gated recurrent unit and a decision maker part composed of a multi-layer perceptron, and decouple the parameters of the feature extractor part and the decision maker part;
[0024] Step 2.3: Set the decoupled feature extractor parameters as shared parameters, and keep them consistent for all clients during the local adaptation phase;
[0025] Step 2.4: Set the decoupled decision maker parameters as personalized parameters and use the local dataset in the local adaptation phase Perform fine-tuning to obtain a personalized model.
[0026] The beneficial effects of the present invention are:
[0027] The present invention adopts a method of global model personalization, and through the technical means of global federated learning plus local adaptation, it not only protects patient privacy information, but also customizes a personalized model for each medical center; through the self-paced learning strategy, according to the order of sample loss values from low to high, the samples are gradually added in the global federated learning process, which effectively solves the problem of limited generalization ability of the global model caused by data heterogeneity and difficulty in converging to the global optimal model. BRIEF DESCRIPTION OF THE DRAWINGS
[0028] Figure 1 It is the overall technical framework of the multi-center data processing method based on self-paced learning and personalized federated learning of the present invention;
[0029] Figure 2 It is a diagram of the parameter update process of the global self-paced learning neural network model;
[0030] Figure 3 Set up the graph for the neural network model. DETAILED DESCRIPTION
[0031] The present invention will be further described below in conjunction with the accompanying drawings and embodiments.
[0032] like Figure 1 As shown, the present invention provides a multi-center data processing method based on self-paced learning and personalized federated learning, and mainly makes technical improvements in the following two aspects:
[0033] First: Global federated learning setting based on self-paced learning.
[0034] In the process of global federated learning, a self-paced learning strategy is used to sort and reselect client training samples, giving priority to learning easier samples in each client, and gradually learning more difficult samples. This method can alleviate the negative impact of heterogeneous data on global federated model training, stabilize the convergence of the global federated model, and enhance the generalization ability of the global federated model.
[0035] By dynamically adjusting the sample selection threshold (from mean loss to maximum loss), it ensures that the client's local model can select suitable samples for learning at different training stages, thereby achieving the best training effect.
[0036] Second: Design and personalized approach of prediction model for ICU patients.
[0037] In order to process the original ICU electronic medical data containing static features, time series features and diagnostic features and perform clinical prediction, the present invention designs a prediction model, which includes a diagnosis encoder, a feature extractor of a two-layer gated recurrent unit and a decision maker composed of a multi-layer perceptron.
[0038] In order to solve the problem of personalized adaptation of the federated model, the present invention divides the prediction model parameters into two parts: one is the global shared parameters trained by global federated learning , and the other is the client-side private parameters trained by the local adaptive method Because in deep learning the model focuses on learning common and low-level features at lower layers, the higher layers learn the specific features of the task, and the lower layers of the model are highly transferable. The present invention sets the feature extractor parameters of the prediction model as shared parameters, while the decision maker parameters are set as private parameters.
[0039] The specific steps that the present invention can achieve are described in detail below:
[0040] Step 1: In the global federated learning phase, the federated model is trained using the global self-paced learning method.
[0041] Step 1.1: The server establishes a connection with the client, determines the global federated learning participants, and initializes the global federated model.
[0042] Step 1.2: The server sends the initialized global federated model to the client.
[0043] Step 1.3: The client accepts the initialized global federated model and uses it to calculate the sample loss value for the local training sample data. .
[0044] Step 1.4: Based on the current number of training rounds t, the total number of training rounds T, and the client local data set The maximum loss value of all training samples And the average loss value Calculate sample selection threshold .
[0045] ,
[0046] Step 1.5: Screen the data and select the sample loss value Not greater than the sample selection threshold The local training sample set constitutes a new dataset Perform local training; (As shown in Table 1, when the threshold is 0.625, the sample set with a loss value less than or equal to 0.625 is selected as the new data set ).
[0047] Table 1
[0048]
[0049] Step 1.6: Upload the trained client local model to the server. The server aggregates all local models to obtain a new global federated model. The model aggregation formula is as follows: is the number of clients, is the weight of the number of the nth client, u and v are the feature extractor parameters and decision maker parameters of the prediction model respectively), and the neural network model parameter update process during the training process is shown in Figure 2.
[0050] ,
[0051] are the feature extractor parameters of the local model of the nth client in the tth training round, are the decision maker parameters of the client local model of the nth client in the tth training round, are the feature extractor parameters of the global federated model in the t+1th training round are the decision maker parameters of the global federated model in the t+1th training round.
[0052] By weighted aggregation of the model parameters of all clients in the current round, new global federated model parameters can be obtained and sent to the clients for training in the next round.
[0053] Step 1.7: Repeat steps 1.2 to 1.6 until the set total number of training rounds T to obtain the final global federated model.
[0054] Step 2: In the local adaptation phase, a personalized model is obtained through the parameter decoupling method.
[0055] Step 2.1: The client accepts the final global federation model.
[0056] Step 2.2: Divide the final global federated model into a feature extractor consisting of a double-layer gated recurrent unit and a diagnostic encoder, and a decision maker consisting of a multi-layer perceptron. The two parts of the global federated model are set as follows: Figure 3 As shown in Figure 2, the global federated model parameters (u, v) are decoupled according to the two parts, and the feature extractor parameters of the global federated model are and the decider parameters Separate.
[0057] Step 2.3: Set the feature extractor parameters of the final global federated model as shared parameters, which are kept consistent across all clients during the local adaptation phase (frozen and unchanged during training).
[0058] Step 2.4: Set the decision maker parameters of the final global federated model to personalized parameters, and use local data in the local adaptation stage Make fine adjustments.
[0059] Through experimental verification of mortality prediction and hospital stay prediction in 8 centers in a real multi-center eICU database, the results of the method of the present invention are compared with those of single local training (Local), the baseline federated learning algorithm FedAvg and other four personalized federated learning algorithms as shown in Tables 2 and 3 below.
[0060] Table 2. Mortality prediction comparison table (AUROC↑)
[0061]
[0062] Table 3. Comparison of hospital stay predictions (MSE↓)
[0063]
[0064] By analyzing the data in Table 2 and Table 3, it can be clearly observed that the method proposed in the present invention has shown excellent performance in both the mortality prediction task and the hospital stay prediction task. Among the 8 participating clients, the present invention won 5 best scores in the mortality prediction task, and even achieved 7 best scores in the hospital stay prediction task. In addition, according to the weighted average calculation of the number of clients, the overall performance of the present invention also ranks first. Compared with separate local training, the benchmark federated learning algorithm, and the suboptimal personalized federated learning algorithm, the performance of the present invention in the mortality prediction task has increased by 4.19%, 3.83%, and 2.29%, respectively; and in the hospital stay prediction task, it has reached 35.51%, 9.12%, and 4.07%, respectively. This achievement fully demonstrates the advanced nature and practical value of the present invention.
[0065] The specific embodiments described above further illustrate the objectives, technical solutions and beneficial effects of the present invention in detail. It should be understood that the above description is only a specific embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the protection scope of the present invention.
Claims
1. A multi-center data processing method based on self-paced learning and personalized federated learning, characterized in that: The method comprises the following steps: Step 1: Train the global federated model based on the global self-paced learning strategy: collect patient-related information as local training samples, sort and select client training samples based on the training sample loss value and the sample selection threshold, and dynamically adjust the sample selection threshold so that the global federated model selects different client training samples at different training stages until the training termination condition is reached to obtain a trained global federated model; the patient-related information includes patient static data, patient diagnosis data, and patient time series data; the patient static data includes patient age, gender, and height; Step 2, build a personalized model for each client: split the trained global federated model into a feature extractor part and a decision maker part, and decouple the parameters of the feature extractor part and the decision maker part, set the decoupled feature extractor parameters as shared parameters, and set the decision maker parameters as client personalized parameters to form a personalized model for each client, and the client predicts the patient's diagnosis results based on its own personalized model.
2. The multi-center data processing method based on self-paced learning and personalized federated learning according to claim 1 is characterized in that: The step 1 comprises: Step 1.1: The server establishes a connection with the client, determines the global federated learning participants, and initializes the global federated model; Step 1.2, the server sends the initialized global federated model to the client; Step 1.3: The client accepts the initialized global federated model and calculates the single sample loss value ℒ for the local training sample data; Step 1.4: Based on the current number of training rounds 𝑡, the total number of training rounds 𝑇, and the client local data set The maximum loss value of all training samples And the average loss value , calculate the sample selection threshold 𝜆; , in, Indicates the client local data set when the current training round number is t The average loss value of all training samples, Indicates the client local data set when the current training round number is t The maximum loss value of all training samples; Step 1.5: Select a local training sample set whose single sample loss value ℒ is not greater than the sample selection threshold 𝜆 to form a new dataset Conduct training; Step 1.6: Based on the new dataset The trained client local model is uploaded to the server, and the server aggregates all local models to obtain a new global federated model; Step 1.7: Repeat steps 1.2 to 1.6 until the set total number of training rounds T is met to obtain a trained global federated model.
3. The multi-center data processing method based on self-paced learning and personalized federated learning according to claim 1 is characterized in that: The step 2 comprises: Step 2.1: The client accepts the final global federation model; Step 2.2, split the final global federated model into a feature extractor including a double-layer gated recurrent unit and a decision maker part composed of a multi-layer perceptron, and decouple the parameters of the feature extractor part and the decision maker part; Step 2.3: Set the decoupled feature extractor parameters as shared parameters, and keep them consistent for all clients during the local adaptation phase; Step 2.4: Set the decoupled decision maker parameters as personalized parameters and use the local dataset in the local adaptation phase Fine-tuning is performed to obtain a personalized model, and the client predicts the patient's diagnosis result based on the respective personalized model.
