A pulse disease diagnosis device fusing a manual feature and a deep feature

Abstract

The application relates to a pulse disease diagnosis device fusing manual features and deep features, which mines information with physiological significance of a pulse signal through a traditional manual feature extraction method, mines deep feature information of the pulse signal through a deep learning method, and fuses the manual features and the deep features for pulse disease diagnosis; compared with the prior art, the diagnosis performance of the diagnosis device is improved.

Description

Technical Field

[0001] This application relates to the field of machine learning technology, and more specifically, to a pulse disease diagnostic device that integrates handcrafted features and deep features. Background Technology

[0002] Pulse diagnosis in Traditional Chinese Medicine (TCM) is one of the most representative diagnostic methods among the four diagnostic methods of TCM: observation, auscultation, inquiry, and palpation. Traditional TCM pulse diagnosis is widely used in clinical practice, and accurate use of the pulses in both hands is particularly important for assessing a patient's physiological and pathological information. TCM practitioners typically use the "three parts and nine pulses" method to diagnose the state and disease information of different organs. Skilled doctors can use pulse information to determine a patient's lifestyle, constitution, and past medical history. However, this diagnostic method is greatly influenced by the subjective and objective factors, and pulse diagnosis lacks quantitative data and objectivity. With the development of sensor and artificial intelligence technologies, computer-aided medicine is gaining widespread application.

[0003] Existing research on the objectification of pulse diagnosis has achieved good results in disease identification, but some shortcomings remain. Since traditional time-domain and frequency-domain features cannot fully describe pulse characteristics, some researchers are currently using deep learning to extract pulse features. Deep networks have excellent feature extraction and information representation capabilities, and can extract deep-level features of pulse signals. However, relying solely on deep learning may overlook empirical features from the field of Traditional Chinese Medicine. While traditional handcrafted features have high specificity and biological significance for different diseases, their waveform local analysis methods are insufficient for analyzing the nonlinear long-term structure of pulse signals. Summary of the Invention

[0004] To overcome at least one deficiency in the prior art, this application provides a pulse disease diagnostic device that integrates manual features and depth features.

[0005] In a first aspect, a pulse disease diagnostic device integrating manual and depth features is provided, comprising:

[0006] The pulse signal acquisition and period division unit is used to acquire the wrist pulse signal of the user to be identified and perform period division to obtain multiple single-cycle pulse segments.

[0007] The manual feature extraction unit is used to extract the temporal, frequency, and morphological features of each single-cycle pulse segment to form manual features;

[0008] The deep feature extraction unit is used to input each single-cycle pulse segment into a one-dimensional convolutional neural network to obtain the first deep feature; and to input each single-cycle pulse segment into a Bi-LSTM network to obtain the second deep feature.

[0009] The feature concatenation unit and the dimensionality reduction unit are used to concatenate the first depth feature and the second depth feature and input them into the fully connected layer to reduce the dimensionality and obtain low-dimensional features.

[0010] The feature fusion unit is used to fuse low-dimensional features with handcrafted features to obtain the complete feature vector corresponding to each single-cycle pulse segment.

[0011] The classification unit is used to input the complete feature vector corresponding to each single-cycle pulse segment into the fully connected layer. The output of the fully connected layer is input into the Sigmoid classification model to obtain the pulse disease diagnosis result of the user to be identified.

[0012] In one embodiment, the time-domain features include the amplitude of the main wave, the amplitude of the pre-diplomatic wave, the amplitude of the diplomatic wave, the elasticity coefficient, the mean, the variance, and the absolute error; the frequency-domain features include the energy features extracted by multi-layer wavelet packet transform; and the morphological features include the height of the first peak, the height of the second peak, the height of the third peak, the height of the first trough, the height of the second trough, the peak occurrence time, the time interval between the peak and the trough, and the pulse cycle length.

[0013] In one embodiment, a one-dimensional convolutional neural network includes two convolutional layers, with the kernel size of each convolutional layer being 3×1.

[0014] In one embodiment, the Bi-LSTM network includes two Bi-LSTM units.

[0015] In one embodiment, the classification unit is further configured to: input the complete feature vector corresponding to each single-cycle pulse segment into a fully connected layer, and input the output of the fully connected layer into a Sigmoid module to obtain the pulse disease diagnosis result corresponding to the single-cycle pulse segment, wherein the pulse disease diagnosis result indicates that the user to be identified is a patient with pulse disease or a normal person; if the probability that the user to be identified is a patient with pulse disease in the pulse disease identification results obtained from multiple single-cycle pulse segments exceeds a set proportion, then the user to be identified is considered to be a patient with pulse disease.

[0016] In one embodiment, the device further includes a Sigmoid classification model training module for training a Sigmoid classification model based on wrist pulse signals collected from normal individuals and users with pulse disorders, thereby obtaining a trained Sigmoid classification model.

[0017] Compared with the prior art, the present application has the following beneficial effects: The pulse disease diagnostic device of the present application mines physiologically meaningful information from pulse signals through traditional manual feature extraction methods, and mines deep feature information from pulse signals through deep learning methods, and fuses manual features and deep features for pulse disease diagnosis; compared with the prior art, the diagnostic performance of the diagnostic device of the present application is greatly improved. Attached Figure Description

[0018] This application can be better understood by referring to the description given below in conjunction with the accompanying drawings, which, together with the detailed description below, are incorporated in and form part of this specification. In the drawings:

[0019] Figure 1 A structural block diagram of a pulse disease diagnostic device integrating manual and depth features according to an embodiment of this application is shown;

[0020] Figure 2 A flowchart illustrating the diagnosis of pulse disorders is shown;

[0021] Figure 3 A schematic diagram of a Bi-LSTM network is shown. Detailed Implementation

[0022] Exemplary embodiments of the present application will be described below with reference to the accompanying drawings. For clarity and brevity, not all features of the actual embodiments are described in the specification. However, it should be understood that many embodiment-specific decisions can be made in the development of any such actual embodiment to achieve the developer’s specific objectives, and these decisions may vary as the embodiments differ.

[0023] It should also be noted that, in order to avoid obscuring this application with unnecessary details, only the device structure closely related to the solution according to this application is shown in the accompanying drawings, while other details that are not closely related to this application are omitted.

[0024] It should be understood that this application is not limited to the described embodiments by virtue of the following description with reference to the accompanying drawings. In this document, embodiments may be combined with each other, features may be substituted or borrowed between different embodiments, and one or more features may be omitted in one embodiment, where feasible.

[0025] This application provides a pulse disease diagnostic device that integrates manual and depth features. Figure 1 A structural block diagram of a pulse disease diagnostic device integrating manual and depth features according to an embodiment of this application is shown. Figure 2 A flowchart illustrating the diagnosis of pulse disorders is shown. See [link / reference]. Figure 1 and Figure 2 The device includes:

[0026] The pulse signal acquisition and period division unit 11 is used to acquire the wrist pulse signal of the user to be identified and perform period division to obtain multiple single-cycle pulse segments.

[0027] Here, a piezoelectric sensor pulse acquisition device can be used to collect the pulse signal of the human wrist, perform preprocessing such as noise reduction on the data, and complete the period division.

[0028] The manual feature extraction unit 12 is used to extract the temporal, frequency, and morphological features of each single-cycle pulse segment to form manual features.

[0029] Specifically, the time-domain features include the amplitude of the main wave, the amplitude of the pre-diphthous wave, the amplitude of the diphthous wave, the elasticity coefficient, the mean, the variance, and the absolute error; the frequency-domain features include the energy features extracted by multi-layer wavelet packet transform; and the morphological features include the height of the first peak, the height of the second peak, the height of the third peak, the height of the first trough, the height of the second trough, the peak occurrence time, the time interval between the peak and the trough, and the pulse cycle length.

[0030] The deep feature extraction unit 13 is used to input each single-cycle pulse segment into a one-dimensional convolutional neural network to obtain the first deep feature; and to input each single-cycle pulse segment into a Bi-LSTM network to obtain the second deep feature.

[0031] Specifically, the one-dimensional convolutional neural network (1D-CNN) includes two convolutional layers. The convolutional kernels of the convolutional layers are 3×1 in size and move along the direction of increasing time series. Each time the convolutional kernel moves by one step, it performs a convolution operation on the data within the kernel to extract information. The input pulse signal vector is 50×1 in size. After passing through the convolutional kernels of the two hidden layers, the output of the hidden layer is 46×64 in size.

[0032] The Bi-LSTM network consists of two Bi-LSTM units, and the output size is 60×64 after passing through the two Bi-LSTM units. Figure 3 A schematic diagram of a Bi-LSTM network is shown.

[0033] The feature splicing unit and dimensionality reduction unit 14 are used to splice the first depth feature and the second depth feature and then input them into the fully connected layer FC to reduce the dimensionality and obtain low-dimensional features.

[0034] Here, the fully connected layer (FC) itself has dimensionality reduction capabilities. Through backpropagation and gradient descent, it learns important features in a large vector and outputs a low-dimensional feature.

[0035] Feature fusion unit 15 is used to fuse low-dimensional features with handcrafted features to obtain the complete feature vector corresponding to each single-cycle pulse segment.

[0036] Classification unit 16 is used to input the complete feature vector corresponding to each single-cycle pulse segment into the fully connected layer. The output of the fully connected layer is input into the Sigmoid classification model to obtain the pulse disease diagnosis result of the user to be identified.

[0037] This embodiment extracts physiologically meaningful information from pulse signals using traditional manual feature extraction methods, and extracts deep feature information from pulse signals using deep learning methods. The manual and deep features are then fused for pulse disease diagnosis. Compared with existing technologies, the diagnostic performance of the diagnostic device in this application is significantly improved.

[0038] In one embodiment, the classification unit is further configured to: input the complete feature vector corresponding to each single-cycle pulse segment into a fully connected layer, and input the output of the fully connected layer into a Sigmoid module to obtain the pulse disease diagnosis result corresponding to the single-cycle pulse segment, wherein the pulse disease diagnosis result indicates that the user to be identified is a patient with pulse disease or a normal person; if the probability that the user to be identified is a patient with pulse disease in the pulse disease identification results obtained from multiple single-cycle pulse segments exceeds a set proportion, then the user to be identified is considered to be a patient with pulse disease.

[0039] In one embodiment, the device further includes a Sigmoid classification model training module for training a Sigmoid classification model based on wrist pulse signals collected from normal individuals and users with pulse disorders, thereby obtaining a trained Sigmoid classification model.

[0040] Here, the wrist pulse signals collected from normal individuals and users with pulse disorders are processed through the pulse signal acquisition and period division unit 11, manual feature extraction unit 12, deep feature extraction unit 13, feature splicing unit and dimensionality reduction unit 14, and feature fusion unit 15 in the aforementioned embodiment to obtain a complete feature vector. The complete feature vector is then input into the Sigmoid classification model for iterative training, ultimately resulting in the trained Sigmoid classification model.

[0041] Experiments were conducted on a hypertension / health dataset. The classification model trained using the device and method described in this application achieved a classification accuracy of 96.5%, while the existing method achieved an F1-Score of 91.3%. Table 1 shows the classification performance of this application. As can be seen from Table 1, this application has a significant improvement in classification performance compared with the prior art.

[0042] Table 1

[0043]

[0044] The above descriptions are merely various embodiments of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A pulse disease diagnostic device integrating manual and depth features, characterized in that, include: The pulse signal acquisition and period division unit is used to acquire the wrist pulse signal of the user to be identified and perform period division to obtain multiple single-cycle pulse segments. The manual feature extraction unit is used to extract the time-domain features, frequency-domain features, and morphological features of each single-cycle pulse segment to form manual features; A deep feature extraction unit is used to input each of the single-cycle pulse segments into a one-dimensional convolutional neural network to obtain a first deep feature; and to input each of the single-cycle pulse segments into a Bi-LSTM network to obtain a second deep feature. The feature concatenation unit and the dimensionality reduction unit are used to concatenate the first depth feature and the second depth feature and input them into the fully connected layer to reduce the dimensionality and obtain low-dimensional features; A feature fusion unit is used to fuse the low-dimensional features with the handcrafted features to obtain a complete feature vector corresponding to each single-cycle pulse segment. The classification unit is used to input the complete feature vector corresponding to each single-cycle pulse segment into the fully connected layer, and the output of the fully connected layer is input into the Sigmoid classification model to obtain the pulse disease diagnosis result of the user to be identified.

2. The apparatus as claimed in claim 1, characterized in that, The time-domain features include the amplitude of the main wave, the amplitude of the pre-diplomatic wave, the amplitude of the diplomatic wave, the elasticity coefficient, the mean, the variance, and the absolute error; the frequency-domain features include the energy features extracted by multi-layer wavelet packet transform; the morphological features include the height of the first peak, the height of the second peak, the height of the third peak, the height of the first trough, the height of the second trough, the peak occurrence time, the time interval between the peak and the trough, and the pulse cycle length.

3. The apparatus as described in claim 1, characterized in that, The one-dimensional convolutional neural network includes two convolutional layers, and the size of the convolutional kernel of the convolutional layer is 3×1.

4. The apparatus as claimed in claim 1, characterized in that, The Bi-LSTM network comprises two Bi-LSTM units.

5. The apparatus as claimed in claim 1, characterized in that, The classification unit is further configured to: input the complete feature vector corresponding to each single-cycle pulse segment into a fully connected layer, and input the output of the fully connected layer into a Sigmoid module to obtain the pulse disease diagnosis result corresponding to the single-cycle pulse segment, wherein the pulse disease diagnosis result indicates that the user to be identified is a patient with pulse disease or a normal person; if the probability that the user to be identified is a patient with pulse disease in the pulse disease identification results obtained from multiple single-cycle pulse segments exceeds a set proportion, then the user to be identified is considered to be a patient with pulse disease.

6. The apparatus as claimed in claim 1, characterized in that, The device also includes a Sigmoid classification model training module, which is used to train the Sigmoid classification model based on the wrist pulse signals of normal people and users with pulse diseases, so as to obtain the trained Sigmoid classification model.

Images