A data processing method for the evaluation results of stroke physical examination based on self-integration

Abstract

The present invention discloses a data processing method for stroke physical examination evaluation results based on self-integration, comprising the following steps: S1. Preprocess the text information of the evaluation results, including the following sub-steps: S11. Word segmentation; S12. Construct a vocabulary: After word segmentation of the physical examination evaluation results of each patient, use the bag-of-words method to form a vocabulary; S13. Embed the obtained vocabulary into a high-dimensional vector; S14. Perform binarization processing on the high-dimensional vector; S2. Construct a self-integration-based model: Input the binarized high-dimensional vector into the BNN network and the XNOR-NET network respectively for training, and fuse the outputs obtained by the two networks. The present invention predicts the hemiplegia stage of stroke based on the physical examination evaluation results of doctors. By introducing model self-integration, it can reduce model loss and improve the prediction accuracy of the model.

Description

Technical Field

[0001] The present invention relates to a data processing method for stroke physical examination assessment results based on self-integration. Background Art

[0002] Stroke, also known as apoplexy or cerebral vascular accident (CVA), is an acute cerebrovascular disease, which is a group of diseases caused by sudden rupture of blood vessels in the brain or blood circulation disorders caused by blood vessel obstruction, resulting in brain tissue damage. Stroke is one of the three major diseases leading to human death. In recent years, the incidence rate has shown an obvious upward trend, while the onset age has shown a downward trend. With the development of modern medicine and the improvement of stroke treatment level, there is a phenomenon of decreasing mortality rate and increasing disability rate, which brings a huge burden to the patient's family and society. Early diagnosis of the patient's condition is particularly crucial, which helps to take timely treatment measures. Clinically, doctors mainly evaluate the hemiplegia stage according to the results of the patient's specialized physical examination, and conduct targeted treatment on the patient according to the diagnosis results.

[0003] Chinese patent application "CN106333682A Early diagnosis method for acute ischemic thalamic stroke based on electroencephalogram nonlinear dynamic characteristics" provides a method for early diagnosis of stroke patients. First, EEG signals are collected, then the obtained data is preprocessed, and then the nonlinear dynamic characteristics of the electroencephalogram signals are extracted and analyzed. MATLAB is used to process and analyze the data. Finally, the obtained results are classified and identified to determine whether the patient has thalamic stroke.

[0004] Chinese patent application "CN111613321A An electrocardiogram stroke auxiliary diagnosis method based on dense convolutional neural network" provides an electrocardiogram stroke auxiliary diagnosis method based on dense convolutional neural network. First, an initial database is established, and the training set and test set are allocated; then the training set is used as the original training data and input into the dense convolutional neural network for parameter training to obtain a basic electrocardiogram stroke diagnosis model; finally, the test set is used as the input data and input into the obtained electrocardiogram stroke diagnosis model for identification to obtain the identification result.

[0005] Existing methods using EEG signals and electrocardiograms only consider the patient's clinical activity data, lacking the use of data on the doctor's physical examination assessment of the patient's physical condition, which affects the accuracy of the results. Summary of the Invention

[0006] The purpose of the present invention is to overcome the deficiencies of the prior art and provide a data processing method for stroke hemiplegia stage prediction based on the physical examination assessment results of doctors. By introducing model self-integration, it can reduce model loss and improve the prediction accuracy of the model.

[0007] The object of the present invention is achieved through the following technical scheme: a method for processing stroke physical examination assessment result data based on self-integration, comprising the following steps:

[0008] S1. Preprocessing the text information of the evaluation results includes the following sub-steps:

[0009] S11. Text segmentation: The evaluation result text information is segmented using the n-gram language model, and the result is recorded as <T1,…,T i ,…,T g >, T i is the i-th word, g is the number of participles;

[0010] S12, constructing a vocabulary: After the physical examination and evaluation results of each patient are segmented, a vocabulary is constructed using the bag-of-words method;

[0011] S13, embedding the obtained vocabulary into a high-dimensional vector;

[0012] S14, binarizing the high-dimensional vector;

[0013] S2. Build a model based on self-integration: input the high-dimensional vector after binarization into the BNN network and XNOR-NET network for training respectively, and fuse the outputs of the two networks.

[0014] Furthermore, the specific implementation method of step S13 is: each word T i is assigned to a random d-dimensional bipolar vector: H Ti ∈{-1,1} [d×1] ;

[0015] The permutation operation ρ is applied to H Tj Apply j times, that is, ρ j (H Tj ), using ρ j (H Tj ) represents word T j The relative position in the n-gram; j (H Tj ) forms a j The vector H Tj The corresponding vector is denoted as m j : All n-grams observed in a patient's physical evaluation are grouped together and the n-gram statistics are embedded into the HD vector h:

[0016]

[0017] where k is the total number of n-grams, and f i is the frequency of the i-th n-gram occurrence.

[0018] The specific implementation method of step S14 is as follows: Use w and b to represent the continuous sentence embedding and binary sentence embedding respectively, L represents the dimension of h, and sign() is the sign function; Set a hard threshold s, and convert the value of b in each dimension to 0 or 1 according to the comparison of the hard threshold, that is, when w (i) is greater than s, b (i) is 1, and when less than s, b (i) is 0, where i = 1, 2, 3,..., L, and the calculation formula is as follows:

[0019]

[0020] b (i) and w (i) represent the continuous sentence embedding and binary sentence embedding of the i-th dimension respectively.

[0021] Furthermore, the specific implementation method of step S2 is as follows:

[0022] S21. Use the BNN network and XNOR-NET network for training respectively. The input is the binarized high-dimensional vector, and the feature data is obtained through training;

[0023] The BNN network includes a sequentially connected convolutional layer, batch normalization layer, activation layer, pooling layer, and fully connected layer. The XNOR-NET network includes a sequentially connected batch normalization layer, binary activation layer, convolutional layer, pooling layer, and fully connected layer;

[0024] In the training stage, first perform two forward operations on each training sample: including one random augmentation transformation and one forward operation of the model; Since the augmentation transformation is random, the results of the two forward operations are different, obtaining the outputs of two different hidden layers, resulting in different outputs of the network;

[0025] The self-ensemble loss function consists of two parts; The first term consists of cross-entropy, which is used to evaluate the error of the labeled data; The second term consists of the mean square error of the results of the two forward operations, which is used to evaluate all the data; Among them, the second term contains a time-varying coefficient, which is used to gradually release the error signal of this term, thereby reducing the uncertainty of the parameters; The self-ensemble loss function is as follows:

[0026]

[0027] where B is the hyperparameter batch, G is the input training set, z i and are the outputs of two trainings of the same model, and y iis the label corresponding to sample i, w(t) is a time-dependent weighting function, and C is the number of different classes.

[0028] S22. Combine the outputs of the two models to generate a mixed sample representation: Combine the outputs obtained from the BNN network and the XNOR-NET to reduce overconfidence and improve model accuracy; represent the output of the i-th sample as x i , randomly mix the outputs x i and x j , as well as the one-hot labels of the two samples and to obtain the mixed sample representation and the label refers to the one-hot label before mixing; α is a random number from Ω to 1.00, and Ω is set above 0.5;

[0029] S23. Calculate the loss function: After the combination, input the mixed sample into the fully connected layers of the two networks for prediction to obtain the prediction result y. Both models use the KL divergence loss as the loss function L KL1 and L kl2 . The KL divergence, i.e., relative entropy, is given by the formula:

[0030]

[0031]

[0032] where are the outputs predicted by the BNN network and the XNOR-NET network respectively, y j is the true value, and G is the size of the training set;

[0033] In addition, use the loss function L SE to minimize the output difference between the two models with the same framework and input. The loss function formula is as follows:

[0034]

[0035] where θ1 and θ2 are the parameter sets of the first model and the second model, f represents the neural network model, represents the neurons randomly sampled in the neural network, and D is the mean squared error;

[0036] The total loss function is expressed as L = L KL1 + L KL2 + λ2L SE , where λ2 is a positive value, and the model parameters are obtained by minimizing the total loss function.

[0037] The beneficial effects of the present invention are as follows: The present invention predicts the stages of hemiplegia in stroke according to the physical examination and evaluation results of doctors. By introducing model self-ensemble, it can reduce model loss and improve the prediction accuracy of the model, which is specifically manifested in the following aspects:

[0038] 1) Add a layer of model self-ensemble to improve the accuracy of the model; the loss function takes into account the differences between two models, which can reduce parameter uncertainty and the loss generated between models;

[0039] 2) Mix the outputs of the two models to reduce overconfidence;

[0040] 3) Use the text information of the physical examination and evaluation results of doctors and perform binarization processing, which can greatly save memory and computing. BRIEF DESCRIPTION OF THE DRAWINGS

[0041] Figure 1 It is a flowchart of a data processing method for physical examination and evaluation results of stroke based on self-ensemble;

[0042] Figure 2 It is a schematic diagram of the BNN network structure;

[0043] Figure 3 It is a schematic diagram of the XNOR-NET network structure. DETAILED DESCRIPTION OF THE INVENTION

[0044] The technical solution of the present invention will be further described below with reference to the drawings.

[0045] As Figure 1 shown, a data processing method for physical examination and evaluation results of stroke based on self-ensemble of the present invention includes the following steps:

[0046] S1. Preprocess the text information of the evaluation results, including the following sub-steps:

[0047] S11. Text segmentation: The source of the evaluation results is the evaluation results of doctors, and there may be inconsistent semantic relationship orders and typos, such as "left hemiplegia. The left upper limb and hand have no autonomous movement and cannot be applied to daily activities. The dependence on daily living ability is obvious. Transfer requires the help of family members". Use the n-gram language model to perform word segmentation on the text information of the evaluation results, and the obtained results are recorded as <T1,..., T i ,..., T g >, T i is the i-th word, and g is the number of segmented words;

[0048] S12. Construct a vocabulary: After segmenting the physical examination assessment results of each patient, use the bag-of-words method (for all training texts, regardless of their order of appearance, only considering each vocabulary that appears in the training text as a separate column feature) to construct a vocabulary;

[0049] S13. Embed the obtained vocabulary into a high-dimensional vector; the specific implementation method is as follows: Each word T i is assigned a random d-dimensional bipolar vector: H Ti ∈{-1,1} [d×1] ;

[0050] Apply the permutation operation ρ to H Tj j times, that is, ρ j (H Tj ), and use ρ j (H Tj ) to represent the relative position of word T j in the n-gram; form a vector corresponding to each word T j from ρ Tj (H j ), denoted as m Tj : j : Group together all the n-grams observed in a patient's physical examination assessment, and embed the n-gram statistical information into the HD vector h:

[0051]

[0052] where k is the total number of n-grams, and f i is the frequency of the i-th n-gram.

[0053] S14. Perform binarization processing on the high-dimensional vector; the specific implementation method is as follows: Use w and b to represent the continuous sentence embedding and binary sentence embedding respectively, L represents the dimension of h, and sign() is the sign function; set a hard threshold s, and convert the value of b in each dimension to 0 or 1 according to the comparison with the hard threshold, that is, when w (i) is greater than s, b (i) is 1, and when it is less than s, b (i) is 0, where i = 1, 2, 3,..., L, and the calculation formula is as follows:

[0054]

[0055] b (i) and w (i) represent the continuous sentence embedding and binary sentence embedding of the i-th dimension respectively.

[0056] S2. Build a self-ensemble based model: To make full use of the binarized HD vectors, the BNN network and the XNOR-NET network are used for training respectively. The binarized high-dimensional vectors are input into the BNN network and the XNOR-NET network for training respectively, and the outputs obtained from the two networks are fused.

[0057] The specific implementation method is as follows:

[0058] S21. Use the BNN network and the XNOR-NET network for training respectively. The input is the binarized high-dimensional vector, and the feature data is obtained through training;

[0059] The BNN network includes a convolutional layer, a batch normalization layer, an activation layer, a pooling layer, and a fully connected layer connected in sequence, as Figure 2 shown; the XNOR-NET network includes a batch normalization layer, a binary activation layer, a convolutional layer, a pooling layer, and a fully connected layer connected in sequence, as Figure 3 shown;

[0060] In the training stage, first perform two forward operations on each training sample: including one random augmentation transformation and one forward operation of the model; since the augmentation transformation is random, the results of the two forward operations are different, obtaining the outputs of two different hidden layers (the middle layer of the fully connected layer), resulting in different outputs of the network;

[0061] The self-ensemble loss function consists of two parts; the first term is composed of cross-entropy, which is used to evaluate the error of the labeled data; the second term is composed of the mean square error of the results of the two forward operations, which is used to evaluate all the data; among them, the second term contains a time-varying coefficient, which is used to gradually release the error signal of this term, thereby reducing the uncertainty of the parameters; the self-ensemble loss function is as follows:

[0062]

[0063] where B is the hyperparameter batch, G is the input training set, z i and are the outputs of two trainings of the same model, y i is the label corresponding to sample i, w(t) is a time-dependent weighting function, C is the number of different classes, which is 3 in this embodiment.

[0064] S22. Fuse the outputs of the two models to generate a mixed sample representation: Fuse the outputs obtained from the BNN network and the XNOR-NET to reduce overconfidence, prevent overfitting of the model, and improve the model accuracy; represent the output of the i-th sample as x i , randomly mix the outputs x i and x j of the i-th and j-th samples, as well as the one-hot labels of the two samples and obtain a mixed sample representation and labels refers to the one - hot label before mixing; α is a random number from Ω to 1.00, and Ω is set above 0.5;

[0065] S23. Calculate the loss function: After fusion, input the mixed sample into the fully - connected layers of the two networks respectively for prediction, and obtain the prediction result y. Both models use the KL - divergence loss as the loss function L KL1 and L kl2 . The KL - divergence, that is, relative entropy, is as follows:

[0066]

[0067]

[0068] where are the outputs predicted by the BNN network and the XNOR - NET network respectively, y j is the true value, and G is the size of the training set;

[0069] In addition, use the loss function L SE to minimize the output difference from two models with the same framework and input. The loss function formula is as follows:

[0070]

[0071] where θ1 and θ2 are the parameter sets of the first model and the second model, f represents the neural network model, represents the neurons randomly sampled in the neural network, and D is the mean squared error;

[0072] The total loss function is expressed as L = L KL1 +L KL2 +λ2L SE , where λ2 is a positive value. Obtain the model parameters by minimizing the total loss function, obtain the final model, and then process the evaluation results of the two network models respectively. Fuse the outputs of the two models as the final staging result.

[0073] Those of ordinary skill in the art will realize that the embodiments described herein are to assist the reader in understanding the principles of the present invention. It should be understood that the scope of protection of the present invention is not limited to such specific statements and embodiments. Those of ordinary skill in the art can make various other specific deformations and combinations without departing from the essence of the present invention according to the technical revelations disclosed in the present invention, and these deformations and combinations are still within the scope of protection of the present invention.

Claims

1. A data processing method for the evaluation results of stroke physical examination based on self-integration, characterized in that, It includes the following steps: S1. Preprocess the text information of the evaluation results, including the following sub-steps: S11. Text segmentation: Segment the text information of the evaluation result using the n-gram language model, and denote the result as <T1, …, T i , …, T g >, where T i is the i-th word, and g is the number of segmented words; S12. Construct a vocabulary: After segmenting the physical examination evaluation results of each patient, use the bag-of-words method to form a vocabulary; S13. Embed the obtained vocabulary into a high-dimensional vector. The specific implementation method is as follows: Each word T i is assigned a random d-dimensional bipolar vector: H Ti ∈{-1, 1} [d×1] ; Apply the permutation operation ρ to H Tj j times, i.e., ρ j (H Tj ), and use ρ j (H Tj ) to represent the relative position of the word T j in the n-gram; form a vector corresponding to each word T j by ρ Tj (H j ), denoted as m Tj : j : Group together all the n-grams observed in the physical examination assessment of a patient, and embed the n-gram statistical information into the HD vector h: where k is the total number of n-grams, and f i is the frequency of occurrence of the i-th n-gram; S14. Binarize the high-dimensional vector. The specific implementation method is as follows: Use w and b to represent the continuous sentence embedding and the binary sentence embedding respectively, L represents the dimension of h, and sign() is the sign function. Set the hard threshold s, and convert the value of b in each dimension to 0 or 1 according to the comparison with the hard threshold, that is, when w (i) is greater than s, b (i) is 1, and when it is less than s, b (i) is 0, where i = 1, 2, 3,..., L, and the calculation formula is as follows: b (i) and w (i) respectively represent the consecutive sentence embedding and binary sentence embedding of the i-th dimension; S2. Construct a self-ensemble model: Input the binarized high-dimensional vectors into the BNN network and the XNOR-NET network respectively for training, and fuse the outputs obtained from the two networks; The specific implementation method is as follows: S21. Use the BNN network and the XNOR-NET network for training respectively. The input is the binarized high-dimensional vector, and the feature data is obtained through training; The BNN network includes a convolutional layer, a batch normalization layer, an activation layer, a pooling layer, and a fully connected layer connected in sequence. The XNOR-NET network includes a batch normalization layer, a binary activation layer, a convolutional layer, a pooling layer, and a fully connected layer connected in sequence; In the training stage, first perform two forward operations on each training sample: including one random augmentation transformation and one forward operation of the model; Since the augmentation transformation is random, the results of the two forward operations are different, obtaining the outputs of two different hidden layers, resulting in different outputs of the network; The self-ensemble loss function consists of two parts; The first term is composed of cross-entropy and is used to evaluate the error of the labeled data; The second term is composed of the mean square error of the results of the two forward operations and is used to evaluate all the data; Among them, the second term contains a time-varying coefficient, which is used to gradually release the error signal of this term, thereby reducing the uncertainty of the parameters; The self-ensemble loss function is as follows: where B is the hyperparameter batch, G is the input training set, z i and are the outputs of two trainings of the same model, y i is the label corresponding to sample i, w(t) is a time-dependent weighting function, and C is the number of different classes; S22. Combine the outputs of the two models to generate a mixed sample representation: Combine the outputs obtained from the BNN network and the XNOR-NET to reduce overconfidence and improve the model accuracy; represent the output of the \(i\)-th sample as \(x\). i Randomly mix the outputs \(x\). i and \(x\). j of the \(i\)-th and \(j\)-th samples, along with the one-hot labels and of the two samples to obtain the mixed sample representation and the label where the one-hot label refers to that before mixing. α is a random number from Ω to 1.00, and Ω is set above 0.5; S23. Calculate the loss function: After fusion, the mixed samples are respectively input into the fully connected layers of the two networks for prediction to obtain the prediction result y. Both models use the KL divergence loss as the loss function L KL1 and L kl2 . The KL divergence, that is, the relative entropy, has the following formula: where are the predicted outputs of the BNN network and the XNOR-NET network respectively, and y j is the true value, and G is the size of the training set; Additionally, use the loss function L SE to minimize the output differences from two models with the same framework and input. The loss function formula is as follows: where θ1 and θ2 are the parameter sets of the first model and the second model, f represents the neural network model, represents the neurons randomly sampled in the neural network, and D is the mean squared error; The total loss function is expressed as L = L KL1 + L KL2 + λ2L SE , where λ2 is a positive value, and the model parameters are obtained by minimizing the total loss function.

Images