CN110200624A - Based on convolutional neural networks-Recognition with Recurrent Neural Network-support vector machines mixed model disease identification algorithm - Google Patents

Abstract

The patent of this invention designs the recognition algorithm based on the hybrid model of convolutional neural network-cyclic neural network-support vector machine in the non-invasive diagnosis system, and obtains the data set through preprocessing methods such as data noise. It is the first time to use convolutional neural network in this field. And the recurrent neural network is used as an end-to-end feature extractor to extract the features of the samples, extract the temporal and spatial features of the data respectively, and finally use the support vector machine to classify the extracted feature data. The algorithm finally achieves high accuracy, high specificity and high sensitivity, and can be widely used in the field of non-invasive detection.

Description

Diseases based on convolutional neural network-recurrent neural network-support vector machine hybrid model disease recognition algorithm

【Technical field】

The invention patent is aimed at the field of non-invasive detection of diseases, especially the field of epilepsy screening.

【Background technique】

Although existing disease screening methods have high detection sensitivity and accuracy, they actually rely on expensive equipment and complicated operations, or cause some irreversible damage to the body. Therefore, there is a need for a low-cost, simple and effective non-invasive surgical method to screen for the disease.

EEG contains a lot of physiological and pathological information, can be measured directly on the human body, is suitable for clinical application, can provide diagnostic basis for some brain diseases, and even become an effective treatment for some brain diseases. In recent years, people pay more and more attention to the study of cognitive function. Effective analysis and evaluation of cognitive function is of great significance to the detection and treatment of cognitive impairment diseases. Epilepsy is a chronic disease in which the abnormal discharge of neurons in the brain leads to partial or complete brain dysfunction. EEG contains a wealth of brain function information and has a high reference value for the diagnosis of epilepsy. In the traditional diagnosis process, doctors need to collect patients' EEG data for one or more days. A large amount of EEG data increases the labor intensity of medical staff and reduces the detection efficiency. Moreover, medical staff may be interfered by subjective factors, and there are problems with different inspection standards. disadvantages. Therefore, the intelligent diagnosis of epilepsy becomes particularly important. At present, scholars at home and abroad have done some research and analysis on the diagnosis of epilepsy, but there are still problems of few classification categories and low classification accuracy.

In view of the above phenomenon, the applicant has designed a disease recognition algorithm of a high-precision convolutional neural network-cyclic neural network-support vector machine hybrid model. This algorithm is aimed at the subsequent data processing and identification process of the non-invasive detection system. Experiments show that the method used in this paper can effectively classify EEG data into healthy period and epileptic seizure period, and the classification accuracy has been greatly improved.

【Content of invention】

Aiming at the defects of the above-mentioned prior art, the patent of the present invention designs a disease recognition algorithm of a convolutional neural network-cyclic neural network-support vector machine hybrid model. The patent of this invention combines machine learning and deep learning for the first time in the field of non-invasive diagnosis of brain detection. The system can very well realize the classification of disease patients and healthy people, and perform high-precision disease detection.

In order to realize the effective diagnosis of diseases, the patent of the present invention proposes a pattern recognition algorithm based on the hybrid model of convolutional neural network-cyclic neural network-support vector machine. The feature extraction process of this algorithm is an end-to-end system, and the recurrent neural network selects the recurrent neural network RNN, and it is characterized in that the method includes the following steps, as shown in Figure 1:

Step 1: Use a portable EEG signal acquisition method to collect EEG data;

Step 2: The collected raw EEG data is input into the convolutional neural network (CNN) to extract spatial features.

Step 3: The collected raw EEG data is input into the recurrent neural network RNN to extract time features.

Step 4: Input the spatial features and temporal features together to the support vector machine (SVM) for final classification to obtain the final diagnosis result.

Described step 1 comprises the following steps:

Step 1.1: Data collection: use a portable EEG signal collection method to collect EEG data;

Step 1.2: Data preprocessing: EEG data filtering, using wavelet for filtering, decomposing and reconstructing EEG signals, and obtaining information in the time domain.

Described step 2 comprises the following steps:

Step 2.1: Input the data into a CNN network consisting of two convolutional layers, two max pooling layers, one feature flattening layer and one fully connected layer;

Step 2.2: Extract the features of the input feature flattening layer as the spatial feature set F1[f ₁₁ ,f ₁₂ ...f _1n ].

Described step 3 comprises the following steps:

Step 3.1: Input the data into the three-layer RNN network;

Step 3.2: Extract the features to be input into the feature flattening layer as the time feature set F2[f ₂₁ , f ₂₂ ...f _2n ].

Described step 4 comprises the following steps:

Step 4.1: Combine the spatial feature F1 and the temporal feature F2 into the support vector machine SVM;

Step 4.2: The final data classification is performed by SVM, and the diagnosis result of disease or health is output.

The present invention uses deep learning to extract deep features in parallel, and then uses traditional machine learning as a strong classifier for final disease diagnosis. The adoption of the invention not only overcomes the defect that the original diagnosis algorithm is easy to overfit based on small samples, but also improves the accuracy rate, sensitivity and specificity to more than 94%.

【Description of drawings】

Figure 1 Algorithm flow chart

Figure 2 Algorithm Implementation Process Diagram

Figure 3 Algorithm training process diagram

【Detailed ways】

Below in conjunction with accompanying drawing, describe the implementation process of the inventive method in detail. It should be emphasized that the following descriptions are only exemplary and not intended to limit the scope of the present invention and its application.

In this patent, the portable EEG signal acquisition method is used to collect EEG data, and then the data analysis method is used to identify and classify the sample data. In this paper, the CNN-RNN-SVM algorithm is used. CNN and RNN are used to extract temporal and spatial features, and then the features are combined and then classified by SVM to obtain high accuracy, sensitivity and specificity results.

Fig. 2 is an algorithm implementation process figure, and the present invention comprises the following steps:

Step 1: Use the portable EEG signal acquisition method to collect EEG to obtain sample data, and perform data preprocessing on the data to obtain the initial historical data sample set D;

Step 2: Use the subset of the initial historical data sample set D as the training set, input the convolutional neural network CNN for training, and save the trained parameters inside the model;

Step 3: Use the subset of the initial historical data sample set D as the training set, input the cyclic neural network RNN for training, and save the trained parameters inside the model;

Step 4: Re-input the training sample data into CNN and RNN, extract the deep features and input them into SVM to train the internal parameters of the model, and save the internal parameters of the model;

Step 5: Input the untrained sample data in the historical data as a test set into CNN and RNN respectively and extract their deep features to obtain spatial features and temporal features. The spatial features and temporal features are input to the support vector machine (SVM) for final classification, and the final diagnosis result is obtained.

First collect the signal and input the EEG data in the non-invasive detection system;

Secondly, perform step 1.2 to filter the collected data, use wavelet to filter, decompose and reconstruct the EEG signal, and obtain time domain information.

The data is then fed into a convolutional neural network and a recurrent neural network. Different from traditional machine learning methods, convolutional neural network and recurrent neural network are an end-to-end system that can automatically extract features without requiring time-consuming and labor-intensive manual feature extraction processes. The neural network adjusts the internal parameters through backpropagation and gradient descent, so that the final result can be optimal.

The training of CNN mainly includes the following steps:

Step 2.1: Input the training data into the CNN network composed of two-layer convolutional network and two-layer fully connected network;

Step 2.2: The network adjusts the internal parameters according to the label value through backpropagation and gradient descent, so that the model can better fit the input and output.

Step 2.3: Save the trained CNN model.

The training of RNN mainly includes the following steps:

Step 3.1: Input the training data into the three-layer RNN network;

Step 3.2: The network adjusts the internal parameters according to the label value through backpropagation and gradient descent, so that the model can better fit the input and output;

Step 3.3: Save the trained RNN model.

When training the SVM, since the CNN and RNN feature extractors have been trained, the samples of the training set are re-input into the two networks, and the features are extracted and then classified to the SVM.

The training of SVM mainly includes the following steps:

Step 4.1: Input the training data into the saved CNN model, and extract the features of the input feature flattening layer as the spatial feature set F1[f ₁₁ ,f ₁₂ …f _1n ];

Step 4.2: Input the training data into the saved RNN model, and extract the features of the input feature flattening layer as the time feature set F2[f ₂₁ ,f ₂₂ …f _2n ];

Step 4.3: Combine spatial features F1 and temporal features F2 and input them into SVM to train SVM parameters.

Finally, when the model is trained, test data is fed into the model to verify its accuracy, sensitivity, and specificity. The final model can be used for actual disease diagnosis after saving.

Step 5.1: Input the untrained data into the saved CNN model, and extract the features of the input feature flattening layer as the spatial feature set F1[f ₁₁ ,f ₁₂ …f _1n ];

Step 5.2: Input the untrained data into the saved RNN model, and extract the features of the input feature flattening layer as the time feature set F2[f ₂₁ ,f ₂₂ …f _2n ];

Step 5.3: Combine spatial features F1 and temporal features F2 and input them into SVM for final prediction. The algorithm training diagram is shown in Figure 3.

The evaluation results of the final model are shown in Table 1, and the accuracy, sensitivity, specificity and F1 score are all higher than 94%. The F1 score is an indicator used in statistics to measure the accuracy of a binary classification model. It takes into account both the accuracy and recall of the classification model. Therefore, considering various indicators, the algorithm in this patent has a good practical prospect and generalization ability.

Table 1 Model evaluation results

It should be noted that the above embodiments are only used to illustrate the application of the patented mixed model algorithm of the present invention in the field of EEG non-invasive diagnosis, rather than limiting the patent of the present invention. Those skilled in the art should understand that the technical solution of this design can be modified or equivalently replaced without departing from the spirit and scope of the technical solution of the present invention, which should be covered by the claims of the present invention.

Claims (8)

1. the present invention is based on the pattern recognition algorithm of convolutional neural network-cyclic neural network-support vector machine hybrid model in the non-invasive detection system, and its characteristic design method comprises the following steps: Step 1: Use the portable EEG signal acquisition method to collect and preprocess the EEG data to obtain the initial historical data sample set D; Step 2: Input the initial historical data sample set D into the convolutional neural network CNN for training and save the model parameters. Step 3: Input the initial historical data sample set D into the recurrent neural network RNN for training and save the model parameters. Step 4: Re-input the training set data into CNN and RNN, extract the deep features and input them into SVM for training; Step 5: Input the untrained sample data into CNN and RNN respectively and extract their deep features to obtain spatial features and temporal features. The spatial features and temporal features are input to the support vector machine (SVM) for final classification, and the final diagnosis result is obtained. 2. The pattern recognition algorithm of the non-invasive diagnosis of disease based on the convolutional neural network-cyclic neural network-support vector machine hybrid model according to claim 1, characterized in that it is used in non-invasive diagnosis for pattern recognition. 3. according to the described pattern recognition algorithm of the non-invasive diagnosis of disease based on convolutional neural network-cyclic neural network-support vector machine hybrid model according to claim 1, it is characterized in that described step comprises the following steps 1 comprises the following steps: Step 1.1: Data collection: use a portable EEG signal collection method to collect data on the EEG; Step 1.2: Data preprocessing: The response data of the sensor is filtered by wavelet to form the initial historical data sample set D. 4. according to claim 1, based on the pattern recognition algorithm of the non-invasive diagnosis of disease based on convolutional neural network-cyclic neural network-support vector machine hybrid model, it is characterized in that said step comprises using wavelet transform denoising in said step 1.2, Decompose and reconstruct the EEG signal to obtain time domain information. 5. according to the pattern recognition algorithm of the non-invasive diagnosis of disease based on convolutional neural network-cyclic neural network-support vector machine hybrid model according to claim 1, it is characterized in that described step 2 comprises the following steps: Step 2.1: Input the training data into the CNN network composed of two-layer convolutional network and two-layer fully connected network for training; Step 2.2: Save the trained CNN model. 6. according to claim 1, based on the disease pattern recognition algorithm of convolutional neural network-cyclic neural network-support vector machine hybrid model, it is characterized in that described step 3 comprises the following steps: Step 3.1: Input the training data into the three-layer RNN network for training; Step 3.2: Save the trained RNN model. 7. The disease non-invasive diagnosis pattern recognition algorithm based on convolutional neural network-cyclic neural network-support vector machine hybrid model according to claim 1, it is characterized in that described step 4 comprises the following steps: Step 4.1: Input the training data into the saved CNN model, and extract the features of the input feature flattening layer as the spatial feature set F1[f11,f12...f1n]; Step 4.2: Input the training data into the saved RNN model, and extract the features of the input feature flattening layer as the time feature set F2[f21,f22...f2n]; Step 4.3: Combine spatial features F1 and temporal features F2 and input them into SVM for parameter training. 8. The disease non-invasive diagnosis pattern recognition algorithm based on convolutional neural network-cyclic neural network-support vector machine hybrid model according to claim 1, it is characterized in that described step 5 comprises the following steps: Step 5.1: Input the untrained data into the saved CNN model, and extract the features of the input feature flattening layer as the spatial feature set F1[f11,f12...f1n]; Step 5.2: Input the untrained data into the saved RNN model, and extract the features of the input feature flattening layer as the time feature set F2[f21,f22...f2n]; Step 5.3: Combine spatial features F1 and temporal features F2 and input them into SVM for final prediction.