A distributed medical auxiliary diagnosis method with privacy protection and smooth correction

Abstract

The application provides a distributed medical auxiliary diagnosis method with privacy protection and smooth correction, comprising the following steps: S1: constructing a connected network composed of multiple hospitals or medical institutions (hereinafter referred to as nodes), the network contains n nodes, only the basic connected condition needs to be met, that is, any node only needs to be connected with part of the nodes, and can communicate with all nodes in the network, and the node does not need to be directly connected with all nodes; S2: collecting m internal characteristic data to form a vector set; S3: establishing a logistic regression problem for the medical diagnosis problem; S4: calling the Exact diffusion distributed optimization algorithm with smooth correction; S5: calculating the value of the logistic function by using the obtained and new patient characteristic data; S6: making a medical diagnosis. Different from other distributed optimization algorithms, the nodes in the Exact diffusion algorithm do not transfer gradient information and have the smooth correction effect with deviation, which brings more strict protection to the privacy data and higher prediction accuracy.

Description

Technical Field

[0001] This invention relates to the field of medical technology, and in particular to a medical auxiliary diagnostic method. Background Technology

[0002] With the continuous development of machine learning and big data technologies, their role as an auxiliary means in medical diagnosis is becoming increasingly important. On the one hand, modern diseases are numerous and complex to distinguish; on the other hand, doctors have heavy workloads and are prone to misdiagnosis when faced with a large number of pathological examinations.

[0003] Currently, the use of machine learning classification methods for medical auxiliary diagnosis (such as the diagnosis of diseases like diabetes and cancer) is under active exploration, and most of the methods developed have the following prominent problems:

[0004] (1) Centralized machine learning methods require the centralized collection of patients' raw data before machine learning and diagnosis. In a single hospital setting, the small amount of patient data used for centralized learning leads to low efficiency and poor accuracy in machine learning methods. Therefore, it is often necessary to conduct joint diagnosis by multiple hospitals using a large number of case samples. In this case, each hospital strictly protects the patient's examination data, and if multiple hospitals need to conduct joint diagnosis, there is a significant risk of privacy leakage.

[0005] (2) Even in a distributed environment, most machine learning methods based on distributed optimization algorithms need to transmit gradient data, and the gradient values ​​are calculated using existing raw data, which are very easy for malicious actors to use for data recovery.

[0006] (3) Insufficient diagnostic accuracy. Medical diagnosis usually requires very high accuracy. When using machine learning methods based on distributed optimization, most algorithms cannot achieve the same level of model learning accuracy as centralized algorithms. Even if a few algorithms achieve relatively accurate model learning, there is still room for improvement in accuracy.

[0007] (4) High computational and communication costs. To obtain accurate solutions, most accurate distributed optimization algorithms rely on frequent exchanges of gradient information, resulting in very high gradient computation and communication costs. Summary of the Invention

[0008] To address the problems existing in the prior art, this invention provides a distributed medical auxiliary diagnosis method with privacy protection and smooth correction to solve them one by one. First, a distributed joint diagnosis mode is proposed, in which each medical institution only needs to cooperate (i.e. communicate) with a few medical institutions, and does not need to cooperate with all medical institutions, thus greatly saving communication costs. Medical diagnosis is made through a limited number of diffusion learnings. Second, an Exact diffusion distributed optimization algorithm with smooth correction is proposed. This algorithm not only inherits the advantages of Exact diffusion, including (1) it does not need to transmit the original data, but only the intermediate parameter values ​​in the iteration process. It does not contain any information of the original data, thus further strengthening the role of privacy protection; (2) it has high learning accuracy and low communication and computation costs. At the same time, our method improves the stability of parameter estimation on the basis of Exact diffusion and avoids the performance degradation of final state estimation.

[0009] The specific technical solution of the present invention includes:

[0010] A distributed medical auxiliary diagnostic method with privacy protection and smooth correction includes the following steps:

[0011] S1: Construct a connected network consisting of multiple hospitals or medical institutions (hereinafter referred to as nodes). This network contains n nodes and only needs to satisfy the basic connectivity condition, that is, any node i only needs to be connected to some nodes (described as N). i Maintaining connectivity allows communication with all nodes within the network, without requiring node i to be directly connected to all nodes. Figure 2 This is a schematic diagram of a network containing 10 medical institutions with an average connectivity of 4.

[0012] S2: Node i collects m internal feature data to form a vector set, described as follows: in Given a p-dimensional set of real numbers and the corresponding label data γ i,l ∈{-1,1}, where 1 represents positive and -1 represents negative.

[0013] S3: Establish the medical diagnosis problem as a binary logistic regression problem, where the logistic function for each node is...

[0014]

[0015] Where μ is the normalization parameter, (·) TLet x be the transpose of the vector, and let x be the regression coefficient or model parameter in a machine learning problem. Our goal is to minimize the above objective function based on a connected network to obtain the optimal x. * ,Right now

[0016]

[0017] S4: Invoke the Exact diffusion distributed optimization algorithm with smooth correction to iteratively calculate x. * A precise approximation;

[0018] S5: Using the obtained x * and new patient characteristic data h new Calculate the value of the logistic function

[0019]

[0020] S6: Based on g(h) new The value of g(h) is used to make a medical diagnosis for new cases, that is, when g(h) new A positive result is diagnosed when g(h) ≥ 0.5. new A value less than 0.5 indicates a negative diagnosis.

[0021] Furthermore, the above-mentioned smoothed correction Exact diffusion distributed optimization algorithm includes the following steps:

[0022] Step 1: Construct a distributed network topology based on the cooperative relationships among multiple medical institutions;

[0023] Step 2: Based on the node connectivity, set the weight coefficient w between nodes. i,j It satisfies the following property: when At that time, w i,j =0; when j∈N i hour, and w i,j ≥0; The typical method for determining the value is the average rule, that is...

[0024]

[0025] In the above formula, n i Let i be the connectivity of node i;

[0026] Step 3: Collect historical diagnostic data of m patients for each node i and record it as a feature vector. l∈{1,2,...,m} and diagnostic label γ i,l ∈{-1,1};

[0027] Step 4: Initialize step size α∈(0,1), normalization parameter μ, smoothing parameter θ∈(0,1), and regression coefficients. and local coefficient estimates Local Deviation Correction Vector And calculate the new weight coefficients.

[0028]

[0029] Step 5: Each node i in the network uses local data h i.l and γ i,l Compute local update

[0030]

[0031] In the above formula, t represents the value of the t-th iteration. Is it the local logistic function in The gradient value at that point, and

[0032]

[0033] Step 6: Local Deviation Correction

[0034]

[0035] Step 7: Pass a temporary vector to each node i in the network. Give its neighbor nodes j∈N i and receive j∈N i temporary vector

[0036] Step 8: Each node i in the network performs consistent aggregation, i.e., calculates...

[0037]

[0038] That is, the estimated value of node i in the t-th iteration;

[0039] Step 9: Each node i in the network performs a smoothing operation for local bias correction, i.e., updates the local correction.

[0040]

[0041] Step 10: If If the condition is met, the iteration stops; otherwise, return to step 5 to continue iterating, where ||·|| represents the Euclidean distance between the vectors, and ε is a predefined precision variation.

[0042] Beneficial effects:

[0043] To enhance the reliability of the experiment, we used the widely recognized LIBSVM real-world diabetes dataset, which contains 768 data samples, each with 8 features. We randomly selected 350 data samples as the training set and the remaining 350 data samples as the test set. These data were evenly distributed among 10 medical institutions, with the following topology: Figure 2 As shown. Therefore, the parameter choices in formulas (1) and (2) are m = 35, n = 10, and p = 8, respectively. Furthermore, we set the normalization parameter μ = 0.001 to ensure f i (x) is a strictly convex function and the step size of the optimization algorithm is α = 0.5.

[0044] This invention compares the convergence performance and prediction accuracy of the proposed smoothed modified Exact diffusion with existing Exact diffusion distributed optimization algorithms. The Exact diffusion algorithm, published in a top journal (Yuan K, Ying B, Zhao X, et al. Exact diffusion for distributed optimization and learning—Part I: Algorithm development. IEEE Transactions on Signal Processing, 2018, 67(3):708-723.), exhibits excellent performance with high estimation accuracy and low communication costs. More importantly, unlike other distributed optimization methods, nodes in the Exact diffusion algorithm do not transmit gradient information, thus providing stricter protection for privacy data. Attached Figure Description

[0045] Figure 1 This is a flowchart illustrating the distributed medical auxiliary diagnosis method with privacy protection and smooth correction of the present invention.

[0046] Figure 2 This is a schematic diagram of a network topology containing 10 medical institutions with an average connectivity of 4.

[0047] Figure 3 The flowchart of the Exact diffusion distributed optimization algorithm with smooth correction proposed in the method of this invention is shown below.

[0048] Figure 4 The algorithm iterates using a model generated from the training dataset to learn a timemap of relative error. The relative error is defined as...

[0049]

[0050] The optimal value x * It was obtained through a centralized gradient descent algorithm with small step sizes and more than 500,000 iterations.

[0051] As can be seen from the figure, the Exact diffusion algorithm with a smoothing coefficient θ has a significant improvement in steady-state accuracy, especially when θ = 0.05, the steady-state accuracy is higher and the performance is smoother. In contrast, the existing Exact diffusion algorithm shows a trend of increasing error in steady-state performance after convergence (the curve in the figure curves upwards).

[0052] Figure 5 It is a time-mapping plot of the relative error of the model learning using the test dataset;

[0053] The results and Figure 4 The results were consistent, further verifying that our invention has superior performance.

[0054] Figure 6 The results show that the diagnostic accuracy of the algorithm based on model-learned parameters during the convergence process improves from approximately 66% initially to approximately 77%. This accuracy can be further improved with larger case data samples and richer feature vectors. Detailed Implementation

[0055] This invention is primarily designed to address the potential privacy issues associated with patient data leakage during collaborative diagnosis across different medical institutions. It presents a privacy-preserving distributed diagnosis method that utilizes the high-precision ExactDiffusion distributed optimization algorithm. During the algorithm iteration process, this algorithm only needs to transmit intermediate estimates of regression coefficients without transmitting the original data or gradient information calculated based on the original data.

[0056] See Figure 1 As shown, it specifically includes:

[0057] Step 1: Construct a connected network consisting of multiple hospitals or medical institutions (hereinafter referred to as nodes). This network contains n nodes and only needs to satisfy the basic connectivity condition, that is, any node i only needs to be connected to some nodes (described as N). i Maintaining connectivity allows communication with all nodes within the network, without requiring node i to be directly connected to all nodes. Figure 2 This is a schematic diagram of a network containing 10 medical institutions with an average connectivity of 4.

[0058] Step 2: Node i collects m internal feature data to form a vector set, described as follows: l∈{1,2,...,m}, where Given a p-dimensional set of real numbers and the corresponding label data γ i,l ∈{-1,1}, where 1 represents positive and -1 represents negative.

[0059] Step 3: Establish the medical diagnosis problem as a binary logistic regression problem, with the logistic function for each node being...

[0060]

[0061] Where μ is the normalization parameter, (·) T Let x be the transpose of the vector, and let x be the regression coefficients, or model parameters in a machine learning problem. Our goal is to minimize the above objective function based on a connected network to obtain the optimal x. * ,Right now

[0062]

[0063] Step 4: Invoke the Exact Diffusion Distributed Optimization Algorithm with Smooth Correction to iteratively calculate x. * The precise approximation (see steps 4 to 10 in the specific implementation).

[0064] Step 5: Use the obtained x * and new patient characteristic data h new Calculate the value of the logistic function

[0065]

[0066] Step 6: Based on g(h) new The value of g(h) is used to make a medical diagnosis for new cases, that is, when g(h) new A positive result is diagnosed when g(h) ≥ 0.5. new A value less than 0.5 indicates a negative diagnosis.

[0067] Based on the existing algorithm, this invention proposes a smooth-corrected Exact diffusion algorithm (see schematic diagram). Figure 3 As shown, steps 4 to 10 in the following steps further improve the estimation accuracy and stability.

[0068] The specific implementation of the method of the present invention is as follows:

[0069] Step 1: Based on the cooperative relationships among multiple medical institutions (hereinafter referred to as nodes), construct a system similar to... Figure 2 A distributed network connectivity graph;

[0070] Step 2: Based on the node connectivity, set the weight coefficient w between nodes. i,jIt satisfies the following property: when At that time, w i,j =0; when j∈N i hour, The typical method for determining values ​​is the average rule, that is...

[0071]

[0072] In the above formula, n i Let be the connectivity of node i.

[0073] Step 3: Collect historical diagnostic data of m patients for each node i and record it as a feature vector. l∈{1,2,...,m} and diagnostic label γ i,l ∈{-1,1};

[0074] Step 4: Initialize step size α∈(0,1), normalization parameter μ, smoothing parameter θ∈(0,1), and regression coefficients. and local coefficient estimates Local Deviation Correction Vector And calculate the new weight coefficients.

[0075]

[0076] Step 5: Each node i in the network uses local data h i.l and γ i,l Compute local update

[0077]

[0078] In the above formula, t represents the t-th iteration. Is it the local logistic function in The gradient value at that point, and

[0079]

[0080] Step 6: Local Deviation Correction

[0081]

[0082] Step 7: Pass a temporary vector to each node i in the network. Give its neighbor nodes j∈N i and receive j∈N i temporary vector

[0083] Step 8: Each node i in the network performs consistent aggregation, i.e., calculates...

[0084]

[0085] That is, the estimated value of node i in the t-th iteration;

[0086] Step 9: Each node i in the network performs a smoothing operation for local bias correction, i.e., updates the local correction.

[0087]

[0088] Step 10: If If the condition is met, the iteration stops; otherwise, return to step 5 to continue iterating, where ||·|| represents the Euclidean distance of the vectors, and ε is a predefined precision variation.

[0089] Step 11: Obtain new patient characteristic data h new ,as well as Calculate the value of the logistic function

[0090]

[0091] Step 12: Based on g(h) new The value of g(h) is used to make a case diagnosis for new cases, that is, when g(h) new A positive result is diagnosed when g(h) ≥ 0.5. new A value less than 0.5 indicates a negative diagnosis.

Claims

1. A distributed medical auxiliary diagnostic method with privacy protection and smooth correction, characterized in that, Includes the following steps: S1: Construct a connected network consisting of n nodes, satisfying only the basic connectivity condition, i.e., any node... It only needs to maintain connectivity with a subset of nodes to communicate with all nodes within the network, without requiring any nodes to be connected. Directly connected to all nodes; S2: Node Collect m internal feature data to form a vector set, described as follows: ,in Given a p-dimensional set of real numbers and corresponding label data. 1 indicates a positive result, and -1 indicates a negative result; S3: Establish the medical diagnosis problem as a binary logistic regression problem, where the logistic function for each node is... , （1） in For normalized parameters, This is the transpose of the vector. These are the regression coefficients, or model parameters in machine learning problems; our goal is to minimize the logistic function based on a connected network to obtain the optimal value. ,Right now ； （2） S4: Call the Exact diffusion distributed optimization algorithm for smoothing deviation correction, and iteratively solve for... A precise approximation; S5: Utilize the obtained and new patient characteristic data Calculate the value of the logistic function （3） S6: Based on The value is used to make a medical diagnosis for new cases, that is, when The diagnosis was positive at that time. The initial diagnosis was negative; in, The Exact diffusion distributed optimization algorithm for smoothing deviation correction includes the following steps: Step 1: Build a distributed network based on partnerships with multiple medical institutions; Step 2: Set the weight coefficients between nodes based on their connectivity. It satisfies the following property: when hour, ;when hour, as well as The typical method for determining values ​​is the average rule, i.e. （4） In the above formula For nodes Connectivity; Step 3: Each node Collect historical diagnostic data from m patients and record them as feature vectors. , and diagnostic labels ; Step 4: Initialize step size Normalization parameters Smoothing parameters regression coefficient and local coefficient estimates Local deviation correction vector And calculate the new weight coefficients. ； （5） Step 5: Each node in the network Based on local data and Compute local update （6） In the above formula Indicates the first The value of the nth iteration. Is it the local logistic function in The gradient value at that point, and ； （7） Step 6: Local Deviation Correction ； （8） Step 7: Each node in the network Passing temporary vectors Give its neighbor nodes and receive temporary vector ; Step 8: Every node in the network Performing consistent aggregation, i.e., computation ， （9） That is, a node In the The estimated value of the next iteration; Step 9: Every node in the network Perform a smoothing operation to correct local bias, i.e., update the local correction. ； （10） Step 10: If If the condition is met, stop iterating; otherwise, return to step 5 and continue iterating. Represents the Euclidean distance between vectors. For predefined precision variations.

Images