A privacy protection medical data analysis method based on blockchain and federated learning

Abstract

This invention relates to the field of medical data analysis, and more particularly to a privacy-preserving medical data analysis method based on blockchain and federated learning. The method includes: collecting and preprocessing raw medical data; obtaining comprehensive features based on the preprocessed raw medical data; training a medical model to obtain prediction results; evaluating and optimizing the medical model, generating updated values ​​for medical model parameters, encrypting and signing these updated values ​​to generate a signature; verifying the encrypted updated values ​​and signature, integrating the encrypted updated values ​​to obtain aggregated updated values, encrypting and decrypting these aggregated updated values ​​to obtain the final updated values; and optimizing the medical model based on the final updated values. This method addresses the problems of federated learning potentially exposing individual privacy during model training; the lack of security and scalability in blockchain; and the high computational and communication overhead of model aggregation when dealing with a large number of participants and massive amounts of medical data, which negatively impact the efficiency and speed of medical data analysis.

Description

Technical Field

[0001] This invention relates to the field of medical data analysis, and in particular to a privacy-preserving medical data analysis method based on blockchain and federated learning. Background Technology

[0002] In the healthcare field, the privacy and security of medical data are paramount. Traditional methods of medical data analysis often involve the centralized storage of sensitive medical data, which can pose risks of data breaches and privacy violations. To address these issues, federated learning and blockchain technologies have emerged and are beginning to play a role in medical data analysis, providing novel solutions for the privacy protection and secure sharing of medical data. Federated learning enables participants to collaboratively build medical models without sharing raw medical data, while blockchain technology ensures the transparency and security of medical data. Combining these two technologies enables the privacy protection and secure analysis of medical data, bringing a more reliable and efficient data management approach to the healthcare industry.

[0003] Existing federated learning and blockchain technologies still face several technical challenges: Regarding healthcare data privacy, while federated learning allows for model training in a distributed environment, individual privacy may still be exposed during model aggregation; regarding secure sharing of healthcare data, although blockchain provides a distributed and tamper-proof method for storing healthcare data, it still lacks security; and regarding the efficiency of healthcare data analysis, although federated learning can train models in a distributed manner, the computational and communication overhead of model aggregation can become extremely high when the number of participants or the volume of healthcare data is large, impacting the efficiency and speed of healthcare data analysis. Therefore, improving the security, efficiency, and accuracy of healthcare data analysis remains a pressing technical challenge. Summary of the Invention

[0004] This invention proposes a privacy-preserving medical data analysis method based on blockchain and federated learning. This method addresses several issues by introducing adaptive feature fusion technology, a dynamic blockchain size adjustment algorithm, and a secure fusion aggregation algorithm. These issues include: in terms of medical data privacy, although federated learning allows model training in a distributed environment, individual privacy may still be exposed during model aggregation; in terms of secure sharing of medical data, while blockchain provides a distributed and immutable method for storing medical data, it still lacks security; and in terms of efficiency in medical data analysis, although federated learning can train models in a distributed manner, the computational and communication overhead of model aggregation can become very high when the number of participants or the amount of medical data is large, affecting the efficiency and speed of medical data analysis.

[0005] A privacy-preserving medical data analysis method based on blockchain and federated learning includes the following steps:

[0006] S1: Utilize a dynamic blockchain resizing algorithm to monitor data transmission load in real time and dynamically adjust block size and shard count; collect and preprocess raw medical data from medical institutions, and obtain comprehensive features based on the preprocessed raw medical data through adaptive feature fusion technology;

[0007] S2: Build and train a medical model using comprehensive features and multi-layer hidden layer operations to obtain prediction results; evaluate and optimize the medical model based on the prediction results, generating updated values ​​for the medical model parameters; encrypt the updated values ​​to generate encrypted updated values; sign the encrypted updated values ​​to generate a signature; verify the encrypted updated values ​​and the signature, and integrate the encrypted updated values ​​to obtain aggregated updated values; encrypt and then decrypt the aggregated updated values ​​to obtain the final updated values; and optimize the medical model based on the final updated values.

[0008] Preferably, S1 specifically includes:

[0009] The data transmission load at time t is calculated using the following formula:

[0010]

[0011] Among them, L t N is the data transmission load at time t. t-k S is the number of encrypted parameter update values ​​transmitted within time tk. t-k The amount of data transmitted is the time tk encrypted parameter update value, F. t-k tk is the transmission frequency of the encryption parameter update value, α, β, and γ are weighting coefficients used to adjust the impact of different load indicators on data transmission load, and v is the size of the sliding window.

[0012] Preferably, S1 specifically includes:

[0013] The raw medical data includes structured medical raw data and image medical raw data; the preprocessed raw medical data includes structured medical data and image medical data; structured medical data features and image medical data features are extracted from the structured medical data and image medical data, respectively.

[0014] Preferably, S1 specifically includes:

[0015] Adaptive feature fusion technology combines structured medical data features and image medical data features, and uses an attention mechanism to assign importance weights to the combined features.

[0016] Preferably, S1 specifically includes:

[0017] The spliced ​​features are weighted and fused based on importance weights to obtain comprehensive features.

[0018] Preferably, S2 specifically includes:

[0019] The medical model performs feature transformation on the comprehensive features through multiple hidden layers and outputs the prediction results using the output layer; the calculation process of the hidden layers is as follows:

[0020] The formula for calculating the first hidden layer is:

[0021]

[0022] Among them, H (1) It is the output of the first hidden layer. It is the weight matrix of the first hidden layer. It is the bias vector of the first hidden layer. It is a comprehensive feature, and ReLU is the activation function;

[0023] The formula for calculating the hidden layers from the second layer to the Xth layer is:

[0024]

[0025] Among them, H (x) H is the output of the x-th hidden layer and the output of the (x-1)-th hidden layer. (x-1) The result after the xth hidden layer It is the weight matrix of the x-th hidden layer. Here, m is the bias vector of the x-th hidden layer, and m is the weight matrix of the x-th hidden layer. The total number of elements in the x-th hidden layer represents the number of weights in the x-th hidden layer. yes The first in Each weight.

[0026] Preferably, S2 specifically includes:

[0027] The formula for calculating the output layer is:

[0028]

[0029] Among them, H (X) It is the output of the Xth hidden layer, where X represents the total number of hidden layers. It is the weight matrix of the output layer. q is the bias vector of the output layer, and q is the weight matrix of the output layer. The total number of elements in yes The p-th weight in This is the prediction result when the input to the medical model is a composite feature.

[0030] Preferably, S2 specifically includes:

[0031] The secure fusion aggregation algorithm integrates the updated encrypted parameters from different medical nodes and introduces a smoothing factor to calculate the aggregated parameter update value; the calculation formula is as follows:

[0032]

[0033] in, This is the updated value of the aggregation parameter, where M is the number of participating medical nodes, and E(ΔW) is the value of the aggregation parameter update. η ) is the updated value of the encrypted parameters of the medical node η. It is the update value of the aggregation parameters in the previous round; ω is the smoothing factor.

[0034] The beneficial effects of the technical solution in the embodiments of the present invention are:

[0035] 1. The constructed medical model can efficiently integrate the features of structured medical data and image medical data, enhancing the predictive ability of the medical model. By weightedly fusing different features through an attention mechanism, the medical model can automatically identify and utilize important features, thereby improving the accuracy and robustness of the prediction results and providing more accurate and reliable support for clinical decision-making.

[0036] 2. Through the secure fusion and aggregation algorithm, the original medical data of each medical institution remains confidential throughout the transmission and aggregation process. The secure fusion and aggregation algorithm ensures that medical data is not leaked or tampered with during transmission, allowing medical institutions to participate in training and optimization without sharing patient medical data, thus fully protecting patient privacy.

[0037] 3. The dynamic blockchain size adjustment algorithm dynamically adjusts the block size and number of shards based on the medical data transmission load, ensuring the efficient operation of the blockchain under high load. The application of sharding technology improves the parallel processing capability of the blockchain, avoids the bottleneck problem of a single chain, and ensures high throughput and low latency medical data processing, making medical data analysis and medical model training more efficient. Attached Figure Description

[0038] Figure 1 This is a flowchart of a privacy-preserving medical data analysis method based on blockchain and federated learning, as described in this invention. Detailed Implementation

[0039] To further illustrate the technical means and effects adopted by the present invention to achieve the intended purpose, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0040] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains.

[0041] The following description, in conjunction with the accompanying drawings, details a specific scheme for a privacy-preserving medical data analysis method based on blockchain and federated learning provided by this invention.

[0042] See attached document Figure 1 The diagram illustrates a flowchart of a privacy-preserving medical data analysis method based on blockchain and federated learning, provided by an embodiment of the present invention. The method includes the following steps:

[0043] S1: Utilize a dynamic blockchain resizing algorithm to monitor data transmission load in real time and dynamically adjust block size and shard count; collect and preprocess raw medical data from medical institutions, and obtain comprehensive features based on the preprocessed raw medical data through adaptive feature fusion technology;

[0044] In a blockchain, a fixed block size determines the amount of data a block can hold, while the number of shards determines the number of blocks that can be processed and transmitted in parallel. When the block size is fixed and the number of shards is limited, a large number of requests to transmit cryptographic parameter update values ​​will slow down data transmission speed and reduce network transmission efficiency.

[0045] A dynamic blockchain resizing algorithm is used to monitor data transmission load in real time. This load includes the number of encrypted parameter updates sent and received by medical nodes, the amount of data transmitted, and the transmission frequency. The block size and number of shards are dynamically adjusted accordingly. When the data transmission load is high, the block size and number of shards are increased to ensure that more encrypted parameter updates can be transmitted and processed in parallel, improving processing speed and network transmission efficiency, reducing latency, and avoiding network congestion. When the data transmission load is low, the block size and number of shards are reduced to decrease resource consumption of the blockchain and federated learning system, optimizing storage and processing efficiency.

[0046] The data transmission load at time t is calculated using the following formula:

[0047]

[0048] Among them, L t N is the data transmission load at time t. t-k S is the number of encrypted parameter update values ​​transmitted within time tk. t-k The amount of data transmitted is the time tk encrypted parameter update value, F. t-k tk is the transmission frequency of the encrypted parameter update value. α, β, and γ are weighting coefficients used to adjust the impact of different load indicators on data transmission load. v is the size of the sliding window used to smooth load calculation. α, β, γ, and v can be designed according to the specific implementation scenario.

[0049] The formula for dynamically adjusting the block size is as follows:

[0050]

[0051] in, This is the initial result for block size at time t, B t In the final result of block size at time t, B t-1 The final result is at time t-1, block size B. min This is the minimum block size, representing the block size under the lowest data transmission load. (B) max This is the maximum block size, representing the block size under the highest data transmission load. ε is an adjustment factor used to control the adjustment speed. L target The target data transfer load represents the ideal data transfer load level that blockchain and federated learning systems hope to achieve. target B min B max ε can be designed specifically according to the specific implementation scenario.

[0052] The formula for dynamically adjusting the number of fragments is as follows:

[0053]

[0054]

[0055] Among them, P t N represents the actual processing capacity at time t. t T is the number of times the encrypted parameter update value is transmitted within time t. t It is the time required to process all data within time t. It is the moving average processing capacity over time t. The value is the actual processing capacity measured at the initial time point t=0, where λ is the moving average coefficient, and U t N is the number of slices at time t. max P is the maximum amount of data that a shard can process. max P represents the blockchain's greatest data processing capability.max , λ, N max It can be configured according to the specific implementation scenario.

[0056] Medical nodes represent the medical institutions that actually provide medical services. As independent participants in the blockchain and federated learning system, they collect raw medical data from within these institutions. This raw medical data includes structured raw medical data from hospital information systems and electronic medical record systems, as well as image-based raw medical data from medical imaging equipment. Structured raw medical data refers to medical data organized according to certain formats and rules, with clearly defined fields and types, facilitating processing and analysis. Common examples of structured raw medical data include patients' electronic medical records and laboratory test results. Image-based raw medical data refers to image data acquired through various medical imaging devices, existing in the form of pictures. Common examples include X-ray films, CT scans, and ultrasound images.

[0057] First, the raw medical data undergoes preprocessing, including data denoising, data standardization, and data format conversion. Structured medical raw data is preprocessed to obtain structured medical data, and image-based medical raw data is preprocessed to obtain image-based medical data. For structured medical data, statistical methods such as calculating the mean, standard deviation, and correlation coefficient are used to extract structured medical data features f. struct The specific statistical methods can be set according to the specific implementation scenario; for medical image data, convolutional neural networks are used to extract the features f of the medical image data. ing The data denoising, data standardization, data format conversion, statistical methods, and convolutional neural networks mentioned are all existing technologies and will not be elaborated here.

[0058] Furthermore, the medical nodes utilize adaptive feature fusion technology to obtain comprehensive features. This technology combines structured medical data features and image-based medical data features, and uses an attention mechanism to assign importance weights to each feature. This maximizes data utilization at the data level, improving the predictive power and robustness of subsequent medical models.

[0059] The comprehensive features are obtained by concatenating structured medical data features and image-based medical data features and then weighting and fusing them using an attention mechanism. These features are used for subsequent medical model training and prediction. The specific process is as follows:

[0060] The medical node first concatenates the structured medical data features and the image medical data features to obtain the concatenated features. The expression is as follows:

[0061]

[0062] Then use the attention mechanism to assign splicing features Importance weight:

[0063]

[0064] Where 'a' is the attention weight vector, representing the concatenated features. It is a set of importance weights for each feature, and softmax is the activation function. It is the attention weight matrix. It is the attention bias vector and the attention weight matrix. and attention bias vector The derivation process involves deep learning technology, including neural network training, backpropagation algorithm and gradient descent optimization. The deep learning technology mentioned is existing technology and will not be elaborated here.

[0065] Based on the calculated attention weight 'a', the splicing features are... The weighted fusion is performed using the following formula:

[0066]

[0067] in, It is a comprehensive feature. It is the first of the attention weight vectors a The importance weights of each feature It is the first in the splicing feature There are n features, where n is the concatenation feature. The number of features in Let the first value of the attention weight vector a be the first value. The importance weights of each feature are used to perform an exponential calculation.

[0068] S2: A medical model is established and trained using comprehensive features and multi-layer hidden layer operations to obtain prediction results; the medical model is evaluated and optimized based on the prediction results, generating updated values ​​for the medical model parameters; the updated values ​​for the medical model parameters are encrypted to generate encrypted updated values; the encrypted updated values ​​for the encrypted updated values ​​are signed to generate a signature; the encrypted updated values ​​for the encrypted updated values ​​and the signature are verified, and the encrypted updated values ​​for the encrypted updated values ​​are integrated to obtain aggregated updated values; the aggregated updated values ​​for the aggregated updated values ​​are encrypted and then decrypted to obtain aggregated updated values; the aggregated updated values ​​for the aggregated updated values ​​are then decrypted to obtain the final updated values; and the medical model is optimized based on the final updated values.

[0069] Medical nodes utilize comprehensive features to build and train medical models. These models analyze input data to provide accurate predictions, which are specific numerical values ​​such as disease risk scores, patient recovery probabilities, or health indicators. These predictions are used to assess patients' health status, assisting doctors in diagnosis and developing personalized treatment plans, thereby improving the quality and efficiency of medical services. The specific process is as follows:

[0070] In the process of building a medical model, medical nodes will integrate features. As input, features are transformed through multiple hidden layers, and the predicted output of the medical model is calculated in the output layer. The number of hidden layers is specified as X, and the value of X can be set according to the specific implementation scenario. The calculation process of the hidden layers is as follows:

[0071] The formula for calculating the first hidden layer is:

[0072]

[0073] Among them, H (1) It is the output of the first hidden layer. The result after the first hidden layer It is the weight matrix of the first hidden layer. It is the bias vector of the first hidden layer, and ReLU is the activation function.

[0074] The formula for calculating the hidden layers from the second layer to the Xth layer is:

[0075]

[0076] Among them, H (x) H is the output of the x-th hidden layer and the output of the (x-1)-th hidden layer. (x-1) The result after the xth hidden layer It is the weight matrix of the x-th hidden layer. It is the bias vector of the x-th hidden layer. It is the exponential sum of all weights in the x-th layer, and m is the weight matrix of the x-th hidden layer. The total number of elements in the x-th hidden layer represents the number of weights in the x-th hidden layer. yes The first in Each weight.

[0077] The formula for calculating the output layer is:

[0078]

[0079] Among them, H (X) It is the output of the Xth hidden layer. It is the weight matrix of the output layer. It is the bias vector of the output layer. It is the exponential sum of all weights in the output layer, and q is the output layer weight matrix. The total number of elements in the matrix represents the number of weights in the output layer. yes The p-th weight in This is the prediction result when the input to the medical model is a composite feature.

[0080] After obtaining prediction results from the medical model, the model needs to be evaluated and optimized. The prediction performance is assessed by calculating the mean squared error loss function. A high mean squared error loss function indicates a large prediction error in the medical model, requiring further optimization. The optimization method uses gradient descent to generate the updated medical model parameter values ​​ΔW for medical node i. i The mean squared error loss function and gradient descent technique are existing technologies and will not be elaborated upon here. After each optimization, the prediction results and the value of the mean squared error loss function are recalculated until the performance of the medical model reaches the expected standard.

[0081] To protect medical data privacy, medical node i updates the medical model parameter value ΔW. i Encryption is performed using the Advanced Encryption Standard (AES) algorithm to generate an updated encryption parameter value E(ΔW). i Next, the Digital Signature Algorithm (DSA) is used to verify E(ΔW). i Perform signature processing to generate signature σ. i Advanced Encryption Standard (AES) and digital signature algorithms are existing technologies and will not be elaborated upon here. Medical node i will use E(ΔW) i ) and σ i Send to the aggregation node via blockchain.

[0082] The aggregation node extracts the encrypted parameter update values ​​and signatures of each medical node from the blockchain, verifies the signature using a digital signature algorithm, and then uses a secure fusion aggregation algorithm to integrate the encrypted parameter update values ​​of each medical node. This secure fusion aggregation algorithm effectively integrates the encrypted parameter update values ​​of different medical nodes to obtain an aggregated parameter update value while protecting the privacy of medical data. Furthermore, each medical node only shares the encrypted parameter update value, ensuring that patient privacy is not compromised.

[0083]

[0084] in, This is the updated value of the aggregation parameter, where M is the number of participating medical nodes, and E(ΔW) is the value of the aggregation parameter update. η ) is the updated value of the encrypted parameters of the medical node η. This is the aggregation parameter update value from the previous round. In the first round of calculating the aggregation parameter update value, since there is no aggregation parameter update value from the previous round, it is necessary to... Set to zero. ω is the smoothing factor used for balancing. and The impact of this improves the stability and accuracy of the medical model, and it can be customized according to the specific implementation scenario.

[0085] Aggregator nodes use Advanced Cryptography Standard (ACS) algorithms to update aggregation parameter values. The encrypted aggregated parameter update value is sent back to each medical node via the blockchain. The medical node receives the encrypted aggregated parameter update value from the blockchain, decrypts it using an Advanced Encryption Standard (AES) algorithm, and then decrypts the aggregated parameter update value again using the AES algorithm to obtain the final parameter update value. Will The optimization of the medical model completes this training cycle. After each training cycle, the loss function value of the medical model is calculated. A high loss function value indicates a large prediction error, requiring continued iterative optimization until the model's performance reaches the expected standard. The expected standard is set specifically based on the implementation scenario. The final medical model possesses higher prediction accuracy and generalization ability, enabling precise analysis and prediction of medical data, providing reliable support for clinical decision-making. By analyzing patients' historical data and current condition, it can provide predictive results, helping doctors to conduct early intervention, develop personalized treatment plans, and evaluate treatment effects, thereby improving the quality and efficiency of medical services.

[0086] In summary, a privacy-preserving medical data analysis method based on blockchain and federated learning has been developed.

[0087] The order of the embodiments is for illustrative purposes only and does not represent the superiority or inferiority of the embodiments. The processes depicted in the drawings do not necessarily require a specific or sequential order to achieve the desired result. In some embodiments, multitasking and parallel processing are possible or may be advantageous.

[0088] The various embodiments in this specification are described in a progressive manner. The same or similar parts between the various embodiments can be referred to each other. Each embodiment focuses on describing the differences from other embodiments.

[0089] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention, and should all be included within the protection scope of the present invention.

Claims

1. A privacy-preserving medical data analysis method based on blockchain and federated learning, characterized in that, Includes the following steps: S1: Utilize the dynamic blockchain resizing algorithm to monitor data transmission load in real time and calculate the data transmission load at time t, using the following formula: wherein L t is the data transmission load at time t, N t-k is the number of transmission of encryption parameter update value within time t-k, S t-k is the amount of data transmission of encryption parameter update value at time t-k, F t-k is the transmission frequency of encryption parameter update value at time t-k, and a, b, g are weight coefficients for adjusting the influence of different load indicators on the data transmission load, and v is the size of the sliding window. It dynamically adjusts the block size and the number of fragments; collects and preprocesses raw medical data from medical institutions; and obtains comprehensive features based on the preprocessed raw medical data through adaptive feature fusion technology. S2: A medical model is established and trained using comprehensive features and multi-layer hidden layer operations to obtain prediction results; the medical model is evaluated and optimized based on the prediction results, generating updated values ​​for the medical model parameters; these updated values ​​are encrypted to generate encrypted parameter update values; the encrypted parameter update values ​​are signed to generate a signature; the encrypted parameter update values ​​and the signature are verified; a secure fusion aggregation algorithm is introduced to integrate the encrypted parameter update values ​​from different medical nodes, and a smoothing factor is introduced to calculate the aggregated parameter update values; the calculation formula is as follows: in, This is the updated value of the aggregation parameter, where M is the number of participating medical nodes, and E(ΔW) is the value of the aggregation parameter update. η ) is the updated value of the encrypted parameters of the medical node η. It represents the updated aggregation parameters from the previous round; ω is the smoothing factor. The aggregated parameter update value is encrypted and then decrypted to obtain the aggregated parameter update value. The aggregated parameter update value is then decrypted to obtain the final parameter update value. The medical model is then optimized based on the final parameter update value.

2. The privacy-preserving medical data analysis method based on blockchain and federated learning according to claim 1, characterized in that, S1 specifically includes: The raw medical data includes structured medical raw data and image medical raw data; the preprocessed raw medical data includes structured medical data and image medical data; structured medical data features and image medical data features are extracted from the structured medical data and image medical data, respectively.

3. The privacy-preserving medical data analysis method based on blockchain and federated learning according to claim 2, characterized in that, S1 specifically includes: Adaptive feature fusion technology combines structured medical data features and image medical data features, and uses an attention mechanism to assign importance weights to the combined features.

4. The privacy-preserving medical data analysis method based on blockchain and federated learning according to claim 3, characterized in that, S1 specifically includes: The spliced ​​features are weighted and fused based on importance weights to obtain comprehensive features.

5. The privacy-preserving medical data analysis method based on blockchain and federated learning according to claim 1, characterized in that, S2 specifically includes: The medical model performs feature transformation on the comprehensive features through multiple hidden layers and outputs the prediction results using the output layer; the calculation process of the hidden layers is as follows: The formula for calculating the first hidden layer is: Among them, H (1) It is the output of the first hidden layer. It is the weight matrix of the first hidden layer. It is the bias vector of the first hidden layer. It is a comprehensive feature, and ReLU is the activation function; The formula for calculating the hidden layers from the second layer to the Xth layer is: Where X represents the total number of hidden layers, H (x) H is the output of the x-th hidden layer and the output of the (x-1)-th hidden layer. (x-1) The result after the xth hidden layer It is the weight matrix of the x-th hidden layer. Here, m is the bias vector of the x-th hidden layer, and m is the weight matrix of the x-th hidden layer. The total number of elements in the x-th hidden layer represents the number of weights in the x-th hidden layer. yes The first in Each weight.

6. The privacy-preserving medical data analysis method based on blockchain and federated learning according to claim 5, characterized in that, S2 specifically includes: The formula for calculating the output layer is: Among them, H (X) It is the output of the Xth hidden layer, where X represents the total number of hidden layers. It is the weight matrix of the output layer. q is the bias vector of the output layer, and q is the weight matrix of the output layer. The total number of elements in the array. yes The p-th weight in This is the prediction result when the input to the medical model is a composite feature.

Images