Pulse signal diagnosis optimization method based on weighted decision tree algorithm

Abstract

The invention relates to the field of disease diagnosis, and aims to solve the problems that in the conventional research, disease classification is directly carried out on the basis of pulse characteristics, and pulse waveform personalized differences generated by individual physiological differences are not considered, so that healthy subjects or patients with certain special physical conditions are often misdiagnosed. The invention provides a pulse signal diagnosis optimization method based on a weighted decision tree algorithm, which comprises the following steps of: extracting each physiological factor influencing pulse image characteristics of a normal person according to factors such as gender, age, BMI index, PWV index, recent diet and exercise condition of a subject, and performing numerical quantization and balance on pulse characteristics through the decision tree algorithm to obtain a pulse signal diagnosis result. The defect that the disease diagnosis accuracy is affected by individual physiological factors is overcome, various feature values balanced by the decision tree weighting algorithm are converted into multi-dimensional feature vector samples, the multi-dimensional feature vector samples are sent to a classifier for disease recognition, and the disease recognition precision can be improved.

Description

Technical Field

[0001] The invention relates to the field of biological signal diagnosis, and in particular to a pulse signal diagnosis optimization method based on a weighted decision tree algorithm. Background Art

[0002] Any active cells, human tissues and organs will produce regular biological signals that are closely related to the state of life. These phenomena are collectively referred to as biological signals. Modern medicine uses some biological signals of the human body as a way to diagnose diseases, such as EEG signals, ECG signals, EMG signals, auscultation signals, pulse signals, etc. These signals reflect information about the internal environment of the human body. Therefore, disease diagnosis through biological signals is a common and effective disease diagnosis method.

[0003] With the development of embedded technology, smart wearable devices have solved many problems that were difficult to solve in the medical field before. Among them, recording biological signals based on sensor devices that fit the human body and using relevant machine learning algorithms to diagnose diseases is the most common form of combining modern medicine with computer technology, and the present invention is aimed at the field of human pulse signal diagnosis. Pulse signals are the most common basis for human biological signal diagnosis in traditional Chinese medicine. Compared with the more common EEG signals and ECG signals in modern medicine, pulse biological signals have the advantages of being more regular and easy to measure. Pulse signals are a relatively simple physiological signal. They have a spatial non-discrete stationary feature, and feature selection is easy. From the time domain, frequency domain and spatial features as the feature values ​​applied to the classification algorithm, a disease recognition system with a very high accuracy can be established. In recent domestic and foreign research, there are already a large number of diagnostic methods based on data preprocessing and feature analysis and extraction of pulse signals, combined with machine learning algorithms for disease monitoring.

[0004] However, existing research cannot eliminate the pulse waveform differences caused by special individuals from a large amount of pulse biosignal data, and therefore often misdiagnoses healthy subjects with certain special physical conditions. Existing research also lacks analysis of the intrinsic connection between the pulse waveform signal characteristics of various types of subjects, and a large number of existing pulse diagnosis systems based on pulse signal feature extraction lack unified standards and diagnostic basis, resulting in low experimental repeatability. Summary of the invention

[0005] In view of the above situation, the present invention proposes a pulse signal diagnosis optimization method based on a weighted decision tree algorithm, which is mainly used to solve the defect of low applicability of the existing intelligent pulse diagnosis system to various groups of people. Through a set of algorithm systems, the various time domain and spatial characteristics of the pulse waveform electrical signal are weighted by decision processing, thereby eliminating the pulse waveform differences caused by physiological factors between different individuals, and scalarizing the various time domain features in the pulse waveform image of the subject. Through a weighted improvement of a decision tree algorithm from machine learning, the applicability and reliability of the existing pulse signal-based pulse diagnosis system for disease diagnosis of various groups of people are improved, and the disease recognition rate is improved, and the standardization of intelligent pulse diagnosis is further completed.

[0006] The pulse signal diagnosis optimization method based on a weighted decision tree algorithm described in the present invention comprises the following steps:

[0007] Step 1: Sample the subject's pulse data by connecting a corresponding electronic device through a sensor of any specification, and convert the bioelectric signal of the wrist pulse into a data set and record it in a computer device that can be used to run a machine learning algorithm.

[0008] Step 2: Preprocess various biological signals and convert them into data sets for feature extraction algorithms. First, preprocess the pulse signal. The raw data of the pulse electrical signal is sent to the computer device and subjected to noise reduction processing to make the electrical signal data present an observable waveform. Then, apply the existing data processing algorithms such as baseline drift removal, filtering, and smoothing to the pulse waveform, so that the pulse presents waveform data that can be observed and feature values ​​extracted. For the existing pulse waveform data set, decompose several groups of feature values ​​and send them to the disease classification algorithm as training set and test set samples.

[0009] Step 3: The extracted data features are sent to the designed weighted decision tree algorithm according to the subject's physiological characteristics that affect the pulse waveform feature values, such as gender, age, BMI index, PWV index, recent diet and exercise status, to balance these affected feature values.

[0010] Step 4: The pulse biosignal feature data set processed by the weighted decision tree is sent to the linear classification method (LM), convolutional neural network (CNN), graph convolutional neural network (GCNN), support vector machine (SVM), support vector machine kernel method (SVM-KM) and other machine learning classification algorithms according to the training set and test set for classification algorithm identification, and cross-validation is performed based on the final classification results to serve as the basis for diagnosing the subject's disease.

[0011] In step 2, taking the human wrist pulse signal as an example, the feature extraction methods corresponding to the specifications of different sensors will be different.Figure 2 The eigenvalue decomposition method will also produce different eigenvalues. The eigenvalue selection of biological signals requires selecting the common points in the periodic waveform images of multiple groups of samples.

[0012] Step 3: Send the extracted data features into the weighted decision tree for weight balancing.

[0013] The decision tree DT weighted algorithm is mainly used to eliminate the pulse image differences caused by physiological differences among subjects.

[0014] The basic idea of ​​DT is to set a standard pulse image (appropriate status, age, BMI index, exercise and living habits) as a reference pulse image, and extract each physiological factor that affects the characteristics of a normal person's pulse image to match it with the standard image. The parameters of the pulse factor are determined by several physiological features, from large to small as a decision tree, from top to bottom nodes, and gradually determined until the characteristics only have pathological factors. First, statistically determine which are physiological factors and which are pathological factors. After excluding physiological factors, the pathological features are classified using a support vector machine. If physiological factors and pathological factors overlap, the algorithm is used to offset and divide them according to the pathological factors. Weights are set on the image to offset the impact of physiological factors on the pulse image. For example, under the same conditions, the pulse strength of men will be 12% higher than that of women. It is necessary to multiply the pulse parameters of women by 1.2 to achieve balance. Before the subject's pulse waveform data enters the classification model, the data is first sent to the decision tree ( Figure 3 As a DT tree model case of this scheme), each eigenvalue is then multiplied by the corresponding weight according to the reference standard to make each feature closer to the eigenvalue of a normal person.

[0015] The weight parameters are based on the sample parameter differences of a specific physiological characteristic category. The standard is the mean and median of the samples. The weights are defined as follows:

[0016] Step 4: Convert the eigenvalues ​​that have been balanced by the DT weighted algorithm into multidimensional feature vector samples and send them into the classifier for disease identification. The classifier can use existing LM, CNN, GCNN, SVM, SVM-KM and other classification algorithms to classify the training set and test set.

[0017] Step 5: Collect and summarize the classification results, analyze the results of the classification algorithm and derive the disease risk. Verify the success rate of the diagnosis system through the confusion matrix and cross-validate the results. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] Figure 1System flow chart of pulse signal diagnosis optimization method based on weighted decision tree algorithm

[0019] Figure 2 This is an example of the decomposition of the pulse biological signal waveform characteristics.

[0020] Figure 3 Example diagram for the weighted decision tree algorithm DETAILED DESCRIPTION

[0021] In order to make the objectives, technical solutions and advantages of the present invention more clearly understood, the specific implementation modes of the present invention are further described in detail below in conjunction with the accompanying drawings.

[0022] Step 1: Sample the subject's pulse data by connecting a corresponding electronic device through a sensor of any specification, and convert the bioelectric signal of the wrist pulse into a data set and record it in a computer device that can be used to run a machine learning algorithm.

[0023] The pulse waveform image and sampling frequency of the pulse bioelectric signal obtained by sensors of different specifications will also be different, so the pulse waveform data using the feature extraction algorithm is expressed as an approximate value. In addition, the ambient temperature and humidity, the height of the subject's seat and the table, the subject's rest time before measurement, the delay of the acquisition instrument, and the signal spectrum amplification factor are all variables that need to be controlled.

[0024] Step 2: Preprocess various biological signals and convert them into data sets for feature extraction algorithms. First, perform data processing algorithms such as baseline drift removal, filtering, and smoothing to make the pulse present waveform data that can be observed and feature values ​​extracted. Figure 2 Several groups of feature values ​​are decomposed in this way and sent to the disease classification algorithm as training set and test set samples.

[0025] The periodicity dataset of pulse biosignal is based on Figure 2 The feature decomposition scheme used for feature extraction, the various explicit biological characteristics of the human wrist pulse signal can be reflected in this figure, and its various characteristics are as follows:

[0026] according to Figure 2The feature value extraction method of the pulse data of the photoelectric sensor is taken as an example. 11 groups of feature values ​​are extracted from the existing pulse waveform data set for disease classification. The symbols of the feature values ​​represent pulse peak (h1), double spiral notch height (h2), pulse cycle time (T / ms), diastolic peak (h2), systolic time (T1 / ms), diastolic time (T2 / ms), systolic index (T1 / T), diastolic index (T2 / T), vascular contraction range (VR), double spiral notch index (h2 / VR), and vascular contraction index (VR / h1).

[0027] Step 3: Send the extracted data features into the weighted decision tree for weight balancing.

[0028] The decision tree DT weighted algorithm is mainly used to eliminate the pulse image differences caused by physiological differences among subjects.

[0029] The basic idea of ​​DT is to set a standard pulse image (appropriate status, age, BMI index, exercise and living habits) as a reference pulse image, and extract each physiological factor that affects the characteristics of a normal person's pulse image to match it with the standard image. The parameters of the pulse factor are determined by several physiological features, from large to small as a decision tree, from top to bottom nodes, and gradually determined until the characteristics only have pathological factors. First, statistically determine which are physiological factors and which are pathological factors. After excluding physiological factors, the pathological features are classified using a support vector machine. If physiological factors and pathological factors overlap, the algorithm is used to offset and divide them according to the pathological factors. Weights are set on the image to offset the impact of physiological factors on the pulse image. For example, under the same conditions, the pulse strength of men will be 12% higher than that of women. It is necessary to multiply the pulse parameters of women by 1.2 to achieve balance. Before the subject's pulse waveform data enters the classification model, the data is first sent to the decision tree ( Figure 3 As a DT tree model case of this scheme), each eigenvalue is then multiplied by the corresponding weight according to the reference standard to make each feature closer to the eigenvalue of a normal person.

[0030] The weight parameters are based on the sample parameter differences of a specific physiological characteristic category. The standard is the mean and median of the samples. The weights are defined as follows:

[0031] Step 4: Convert the eigenvalues ​​that have been balanced by the DT weighted algorithm into multidimensional feature vector samples and send them into the classifier for disease identification. The classifier can use existing LM, CNN, GCNN, SVM, SVM-KM and other classification algorithms to classify the training set and test set.

[0032] Step 5: Collect and summarize the classification results, analyze the results of the classification algorithm and derive the disease risk. Verify the success rate of the diagnosis system through the confusion matrix and cross-validate the results.

[0033] The above content describes the specific implementation mode of the present invention in combination with the accompanying drawings, and the content does not limit the protection scope of the present invention. On the basis of the technical solution of the present invention, various modifications and different forms of reuse that can be made by technical personnel in this field without creative work are still within the protection scope of the present invention.

Claims

1. The pulse signal diagnosis optimization method based on a weighted decision tree algorithm of the present invention is characterized by: Step 1: Sample the subject's pulse data by connecting a sensor of any specification to a corresponding electronic device, and convert the bioelectric signal of the wrist pulse into a data set and record it in a computer device that can be used to run a machine learning algorithm; Step 2: Preprocess various biological signals and convert them into data sets for feature extraction. First, preprocess the pulse signal. The raw data of the pulse electrical signal is sent to the computer device and subjected to noise reduction processing to make the electrical signal data present an observable waveform. Then, the existing data processing algorithms such as baseline drift removal, filtering, and smoothing are applied to the pulse waveform to make the pulse present waveform data that can be observed and feature values ​​extracted. The existing pulse waveform data set is decomposed into several groups of feature values ​​and sent to the disease classification algorithm as training set and test set samples. Step 3: The extracted data features are sent to the designed weighted decision tree algorithm according to the subject's physiological characteristics that affect the pulse waveform feature values, such as gender, age, BMI index, PWV index, recent diet and exercise status, to balance these affected feature values; Step 4: The pulse biosignal feature data set processed by the weighted decision tree is sent to the linear classification method (LM), convolutional neural network (CNN), graph convolutional neural network (GCNN), support vector machine (SVM), support vector machine kernel method (SVM-KM) and other machine learning classification algorithms according to the training set and test set for classification algorithm identification, and cross-validation is performed based on the final classification results to serve as the basis for diagnosing the subject's disease.

2. Step 2 according to claim 1, characterized in that: The pulse signal is filtered, baseline drift is removed and smoothed, and the processed pulse signal is divided into single cycles; for each single cycle pulse signal, 11 groups of characteristic values ​​are decomposed, including pulse peak (h1), double spiral notch height (h2), pulse cycle time (T / ms), diastolic peak (h2), systolic time (T1 / ms), diastolic time (T2 / ms), systolic index (T1 / T), diastolic index (T2 / T), vascular contraction range (VR), double spiral notch index (h2 / VR), and vascular contraction index (VR / h1).

3. Step 3 according to claim 1, characterized in that: The extracted data features are sent to the designed weighted decision tree algorithm according to the subject's physiological characteristics that affect the pulse waveform feature values, such as gender, age, BMI index, PWV index, recent diet and exercise status, to balance these affected feature values. Based on the differences in pulse characteristics caused by specific physiological characteristics, the definition of the single weight of the decision tree DT (Decision Tree) is as follows:

4. Step 4 according to claim 1, characterized in that: The pulse biosignal feature data set processed by the weighted decision tree is sent to the machine learning algorithm for classification model training according to the training set and the test set, such as linear classification method (LM), convolutional neural network (CNN), graph convolutional neural network (GCNN), support vector machine (SVM), support vector machine kernel method (SVM-KM) and other machine learning algorithms. The cross-validation of each classification result and the confusion matrix is ​​used as the basis for the diagnostic standard.

Images