CN110946556B - Parkinson's resting-state tremor assessment method based on wearable somatosensory network - Google Patents

Abstract

The invention discloses a Parkinson's resting state tremor evaluation method based on a wearable somatosensory network, belonging to the field of wireless sensor networks and data analysis thereof, and particularly relates to a wearable somatosensory network-based method for obtaining the tremor state of Parkinson's patients' arms and identification. The invention measures the posture angles of the upper arm, the lower arm and the wrist, calculates the angle change of the elbow joint and the wrist joint, extracts the characteristics of the angle change, extracts the real-time characteristics of the electromyographic signal, and trains the hidden Marker according to the characteristic data and the UPDRS scale. Kov model, which outputs the current optimal state sequence. The method can provide technical support for the assessment of the degree of arm tremor in Parkinson's patients, and provide theoretical basis for Parkinson's patients, the elderly, the infirm and other people who need to know the occurrence of early Parkinson's disease in time.

Description

Parkinson's resting-state tremor assessment method based on wearable somatosensory network

technical field

The invention relates to the field of wireless sensor networks and data analysis thereof, in particular to a Parkinson's resting state tremor evaluation method based on a wearable somatosensory network.

Background technique

Parkinson's disease is a degenerative disease of the central nervous system characterized by varying degrees of impact on the ability of the body to move. It is more common in older adults, with most cases occurring after age 50. Parkinson's disease primarily affects four areas of motor function, including bradykinesia, resting tremor, limb stiffness, and postural instability. Resting tremor is more prominent in the upper limbs, and the tremor will gradually spread from unilateral to bilateral, and the condition will gradually worsen. Therefore, it is very necessary to identify resting tremor earlier in daily life.

The application of wearable sensors in early tremor assessment is currently a research hotspot in academia and industry. Wearable sensors can be used for high-precision tracking of the human body, long-term physiological signal monitoring, etc. This non-implantable monitoring has been used in the evaluation of clinical patient movement abnormalities. However, most of the existing technical solutions use accelerometers and gyroscopes to directly collect tremor signals. The disadvantage of this solution is that it is easy to mix motion components in daily activities, and the accumulated errors of sensors and environmental noise also make the measured signals unreliable. In addition, in the early stage of the disease, the tremor of the limbs is very slight, and it is difficult to extract its features from the signals collected by the accelerometer and gyroscope, and it is easy to be mistaken as a noise signal and discarded. Therefore, adding the muscle electrical signal when tremor occurs into the evaluation system is a better solution for early symptoms. The surface EMG sensor deployed directly on the arm can output real-time EMG signals, and can identify the early tremor state by analyzing the characteristics of the EMG signals.

Invention patent CN201410381634.6 discloses a wearable system for quantitative detection of the main symptoms of Parkinson's patients. The invention obtains the acceleration and angular acceleration of fingertips and wrists through wearable smart gloves, and describes the degree of tremor and In the state of muscle stiffness, due to the direct use of inertial data, it is easily disturbed by the patient's own posture and daily movements, thus affecting the recognition effect of the system.

Invention patent CN201410833652.3 discloses a method for quantifying tremor symptoms in Parkinson's patients based on approximate entropy and mutual approximate entropy, including thumb tremor data collection, index finger tremor data collection and unified Parkinson's disease scoring scale UPDRS scoring, tremor under specified action The approximate entropy between data approximate entropy and tremor data is calculated, but this method does not add EMG signal features, and the recognition effect of early mild tremor is weak.

SUMMARY OF THE INVENTION

Purpose of the invention: The present invention proposes a resting state tremor assessment method for Parkinson's patients based on a wearable somatosensory network, which realizes accurate identification of the degree of arm tremor under daily behavior, and solves the following technical problems: the daily behavior of patients affects the state of arm tremor. The recognition result of the system reduces the positive detection rate of the system; the degree of early tremor is relatively slight, which is easily mistaken for environmental noise and cannot be accurately recognized.

Technical solution: In order to achieve the purpose of the present invention, the technical solution adopted in the present invention is: calculate the angular variation of the wrist joint and the elbow joint through the attitude data collected by the sensor, and extract the time domain and frequency domain features of the variation, using It is used to characterize different degrees of tremor grades; by extracting the surface EMG signal deployed at the upper arm biceps, the signal characteristics are calculated to characterize different degrees of tremor grades; combined with the UPDRS rating scale, by combining angle characteristics and muscle Electrical characteristics, design a multi-hidden state transition HMM model for tremor state assessment under daily behavior. The method for evaluating the resting state tremor of Parkinson's patient's arm based on the wearable somatosensory network of the present invention specifically includes:

Step 1, deploy sensors on human hands and arms, initialize the sensors, eliminate sensor drift, perform initial calibration, and set sensor sampling frequencies; collect human body posture signals and human body surface EMG signals through the deployed sensors;

In step 2, in the daily movements of patients with Parkinson's disease, the low-frequency components (frequency less than 3 Hz) are derived from the patient's conscious movements, while the high-frequency components include tremor and noise during inter-movement; therefore, according to the resting-state tremor frequency response feature, the attitude signal and the EMG signal are filtered through the band-pass filter respectively, and the filtered attitude signal and EMG signal are obtained. The filtered signal retains the attitude component and the EMG signal component caused by tremor; according to the attitude signal Calculate the attitude angle change;

中；由于采集的前期和后期容易受测试准备和测试停止状态转换的影响，因此按时间轴分别剔除数据集

和

的前后p％的数据，保留按时间轴p％至1-p％的数据，更新数据集

和

优选的，p％取值为10％；Step 3, save the discrete data of the attitude angle change and the discrete data of the EMG signal output in the step 2 in the data set respectively.

Medium; since the early and late stages of acquisition are susceptible to test preparation and test stop state transitions, the datasets are separately culled by time axis

and

p% of data before and after, keep data from p% to 1-p% by time axis, update the dataset

and

Preferably, p% is 10%;

和

分别进行等步长窗口化处理，将处理后的数据集

和

组合，选取其中q％作为训练集，用于HMM模型参数学习，其余的1-q％作为测试集，用于HMM模型识别；Step 4, the dataset

and

Perform equal-step windowing processing respectively, and the processed data set

and

Combination, select q% as the training set for HMM model parameter learning, and the remaining 1-q% as the test set for HMM model recognition;

和

的步长设置为100ms，窗口时长为2s，数据交叠率为95％；q％取值为70％；Preferably, in order to continuously describe the changes in tremor, the data set is

and

The step size is set to 100ms, the window duration is 2s, the data overlap rate is 95%; the q% value is 70%;

Step 5, extracting the time domain and frequency domain features of the attitude angle change of Parkinson's tremor in the resting state and the EMG signal features from the windowed data in the training set and the test set respectively;

Step 6: Define the hidden state set of the HMM model according to the UPDRS Parkinson's Assessment Scale; set state labels for each data segment according to the features extracted from the training set;

Step 7: Initialize the parameters of the HMM model, take the feature parameters extracted from the training set as the training observation sequence, perform model training, calculate the optimal parameters of the HMM model, and obtain a trained HMM model; the features include angle change features and electromyographic signal features ;

Step 8, take the feature parameters extracted from the test set as the test observation sequence, input it into the trained HMM model, solve the hidden state sequence when the probability of occurrence of the observation sequence is the largest, output the current optimal hidden state, and obtain the tremor state evaluation result.

Further, in the step 1, IMU sensors are respectively deployed on the upper arm, lower wall and hand, and an EMG sensor is deployed on the upper arm at the same time; the human body posture data is collected through the IMU sensor, and the electrical muscle signal on the surface of the human body is collected through the EMG sensor; the IMU sensor contains Three-axis accelerometer, three-axis gyroscope and three-axis magnetometer, the attitude signal collected by the IMU sensor is output in the form of quaternion. Preferably, the sampling frequency of the IMU sensor is set to 100 Hz, and the sampling frequency of the EMG sensor is set to 50 Hz. Preferably, the nine-axis IMU sensor with the model GY953 is selected, and the EMG sensor model is OYmotion.

Further, in step 2, the upper and lower cut-off frequencies of the band-pass filter through which the attitude signal passes are 3 Hz and 6 Hz, and the upper and lower cut-off frequencies of the band-pass filter through which the myoelectric signal passes are 3 Hz and 12 Hz.

Further, in the step 2, the current coordinate system is set as the earth coordinate system x, y, z, and the attitude angle change is calculated, and the method is as follows:

Define attitude angles θ ₁ , θ ₂ , θ ₃ , attitude angle changes Δθ ₁ , Δθ ₂ , Δθ ₃ , where θ ₁ , θ ₂ , θ ₃ are elbow flexion-extension angle, lower arm internal rotation, respectively -External rotation angle, wrist flexion-extension angle, △θ ₁ , △θ ₂ , △θ ₃ are the change in elbow joint flexion-extension angle, the lower arm inward-external rotation angle change, and the wrist joint flexion-extension angle change quantity;

According to the obtained filtered attitude signal, that is, the attitude quaternion of the upper arm, lower arm and hand IMU sensor, respectively calculate the projection α _x of the attitude angle of the upper arm, lower arm and hand IMU sensor on the coordinate system x, y and z axes, α _y , α _z , β _x , β _y , β _z , γ _x , γ _y , γ _z ;

From the arm geometry we get:

△θ _1x =△α _x +△β _x △θ _3x =△β _x +△γ _x

△θ _1y =△α _y +△β _y , △θ _3y =△β _y +△γ _y

△θ _1z =△α _z +△β _z △θ _3z =△β _z +△γ _z

Among them, △θ _1x , △θ _1y , △θ _1z , △θ _3x , △θ _3y , and △θ _3z are the elbow flexion-extension angle variation Δθ ₁ and the wrist flexion-extension angle variation Δθ ₃ respectively. The projection of the x, y, z axes of the coordinate system; △α _x , △α _y , △α _z is the angle change of the projection of the upper arm attitude angle on the x, y, z axes; △β _x , △β _y , △β _z is the angular change of the projection of the lower arm attitude angle on the x, y, z axis; △γ _x , △γ _y , △γ _z is the angular change of the projection of the hand attitude angle on the x, y, z axis;

Calculate the attitude angle variation Δθ ₁ , Δθ ₃ respectively, the expressions are as follows:

Since the internal rotation or external rotation of the lower arm is rotated along the axis of the elbow joint and the wrist joint, the attitude angle change _Δθ2 is directly derived from the quaternion output by the lower arm IMU sensor. According to the conversion relationship between the quaternion and the attitude angle, the output quaternion can be converted into the angle change.

Further, in the step 5, extract the time domain and frequency domain features of the attitude angle change of Parkinson's tremor in the resting state to describe the degree of tremor, as follows:

描述震颤程度，分别计算姿态角θ₁,θ₂,θ₃的平均角变化率

单位为度/秒；计算公式如下：Feature 1: The angular change of Parkinson's tremor in the resting state is more significant. Therefore, the average angular change rate of Parkinson's tremor in the time domain of the angular change of the resting state is used.

Describe the degree of tremor and calculate the average angular change rate of attitude angles θ ₁ , θ ₂ , θ ₃ respectively

The unit is degrees per second; the calculation formula is as follows:

where n represents the data length, n=t*f _s , t is the window duration, and f _s is the sampling frequency of the IMU sensor;

Feature 2: Resting tremor in Parkinson's disease is often manifested as involuntary movements of the upper limbs of the body in a certain frequency range. Therefore, the energy of Parkinson's tremor in the frequency domain of angular changes in the resting state is used as a feature to describe the degree of tremor. ;

The attitude signal data in the training set is subjected to discrete Fourier transform, converted into frequency domain data, and the energy value is calculated:

where E is the energy value, f is the frequency point, P(f) is the signal energy spectral density, the unit is dB/sec; a, b are the frequency range boundary values of the resting tremor, respectively; [a, b] is the value of The typical frequency range of resting tremor is 3-6 Hz;

Feature 3: Information entropy describes the degree of irregularity or confusion of information. From the characteristics of the attitude signal spectrum, it can be seen that the information entropy in the frequency domain has a good degree of discrimination. Therefore, the information of the angle change of Parkinson's tremor in the resting state is used. Entropy is used as a feature to describe the degree of tremor; the information entropy T is defined as follows:

为频点f对应的出现概率，对数log取2为底。in

is the occurrence probability corresponding to the frequency point f, and the logarithm log takes 2 as the base.

Further, in the step 5, extract the characteristics of the electromyographic signal to describe Parkinson's tremor, as follows:

Feature 4: The EMG signal of Parkinson's tremor shows steep peaks in some areas when tremor occurs. Therefore, the EMG signal kurtosis coefficient is selected as the feature to describe Parkinson's tremor; the kurtosis coefficient k is defined as follows:

where μ is the data mean, Λ is the mathematical expectation, σ is the standard deviation, and x is the EMG discrete data;

即为过零率。所述归一化方法为：将肌电数据映射至[-1,1]区间，得到新的数据为：x′＝(x-x_mean)/(x_max-x_min)，其中x_mean为数据均值，x_max为最大值，x_min为最小值。Feature 5: The EMG signal of Parkinson's tremor also shows dense oscillations in some areas when tremor occurs, that is, the signal has a high density of fluctuations. Therefore, the zero-crossing rate of EMG signal is selected as the feature to describe Parkinson's tremor; Standard normalization, measure the number of zeros in the normalized EMG data

is the zero-crossing rate. The normalization method is: map the EMG data to the [-1,1] interval, and obtain new data as: x′=(xx _mean )/(x _max -x _min ), where x _mean is the data mean , x _max is the maximum value, and x _min is the minimum value.

Further, in the step 6, according to the UPDRS Parkinson's rating scale, define the hidden state set of the HMM model, set the tremor amplitude to be h, and the hidden state set includes:

state 0: no tremor;

State 1: slight tremor, that is, tremor occurs and 0<h≤1cm;

State 2: Moderate tremor, that is, tremor occurs and 1<h≤3cm;

State 3: severe tremor, that is, tremor occurs and 3<h≤10cm;

State 4: Severe tremor, that is, tremor occurs and h>10cm;

According to the features extracted from the training set, state labels 0, 1, 2, 3, and 4 are set for each data segment respectively.

Further, in the step 7, the HMM model parameters include the HMM hidden state number M, the initial probability π, and the state observation probability B; the initial HMM hidden state number M=5, the initial probability π=[1,0,0, 0,0], the state observation probability B is represented by a set of mixed Gaussian densities; let the parameter λ=(A, B, π), A represents the implicit state transition probability;

Taking the feature parameters extracted from the training set as the training observation sequence, the Baum-Welch algorithm is used for model training, and the optimal parameter λ of the HMM model is iteratively calculated.

Further, in the step 8, the feature parameter Λ′ extracted from the test set is used as the test observation sequence, input into the trained HMM model, and the Viterbi algorithm is used to solve the hidden state sequence when P(Λ′|λ) is the largest, Output the current optimal hidden state (ie 0, 1, 2, 3, 4); among them, the feature parameter Λ' is the set of angle features and EMG features; P(Λ'|λ) means that the Viterbi algorithm solves the given HMM model When the parameter λ is set, the probability of occurrence of the observed sequence Λ'.

Beneficial effects: compared with the prior art, the technical solution of the present invention has the following beneficial technical effects:

The method of the invention can provide technical support for evaluating the degree of arm tremor of Parkinson's patients, and provide theoretical basis for Parkinson's patients, the elderly, the infirm and other people who need to know the occurrence of early Parkinson's disease in time. Since this method adopts the relative angle change characteristics of the human wrist and elbow joints, it can improve the defect of the conventional accelerometer or gyroscope scheme that the motion data is affected by the posture and behavior of the human body, and can better characterize the tremor under daily behavior. state, with better robustness and reliability. In addition, due to the real-time EMG signal characteristics of the EMG sensor deployed at the biceps, the degree of tremor in early Parkinson's disease can be extracted, avoiding the defect that the IMU sensor is difficult to obtain early tremor. Experiments show that under the premise that the tremor level is set to 5 according to the evaluation scale, the recognition rate of the HMM model can reach more than 98%, which has good promotion ability and engineering application value.

Description of drawings

Figure 1 is the dynamic model of the human upper limb kinematic chain;

Fig. 2 is the flow chart of the tremor assessment method based on angle change characteristics and electromyography characteristics;

Figure 3 is a schematic diagram of the tremor level implicit state transition based on the HMM model;

Fig. 4 is the experimental result diagram of the method of the present invention;

DESCRIPTION OF REFERENCE NUMERALS: 1- Upper arm IMU sensor deployment position, 2-EMG sensor measurement electrode deployment position, 3- Lower arm IMU sensor deployment position, 4- Hand IMU sensor deployment position, 5- Elbow flexion-extension angle, 6 - Lower arm internal rotation - external rotation angle, 7 - hand flexion - extension angle.

Detailed ways

The technical solutions of the present invention will be further described below with reference to the accompanying drawings and embodiments. In order to better describe the degree of tremor in daily activities of Parkinson's patients, in the specific implementation process, the present invention defines the following 8 scenarios:

S1: lie flat on the bed;

S2: Sit quietly on a chair and relax your arms naturally;

S3: Sit quietly on a chair with your arms flat on the table;

S4: Stand still, with arms naturally relaxed;

S5: After walking slowly for 10 seconds, sit still on the chair and relax the arms naturally;

S6: After grabbing the tea cup to drink water, place the arm flat on the table;

S7: After grabbing the teacup to drink water, the elbow joint is supported on the table;

S8: Standing state, after bending and stretching the upper arm for 5 times, the arm is naturally relaxed;

The steps of the method of the present invention are respectively performed on the above 8 scenarios, and tremor evaluation models under different scenarios are established to test the reliability and effectiveness of the method. Taking the S3 scenario as an example, the implementation process is as follows:

Step 1: Deploy three nine-axis IMU sensors with model GY953 on the upper arm, lower wall and hand of the human body respectively, and deploy the EMG sensor with model OYmotion on the biceps of the arm, as shown in Figure 1; Initialize the sensor, eliminate sensor drift, and perform initial calibration, set the sampling frequency of the IMU sensor to 100Hz, and the sampling frequency of the EMG sensor to 50Hz.

The human body attitude signal is collected by the deployed three IMU sensors, and the electromyographic signal of the human body surface is collected by the EMG sensor; the attitude signal collected by the IMU sensor is output in the form of quaternion.

Step 2: Filter the attitude signal through band-pass filters with upper and lower cut-off frequencies of 3 Hz and 6 Hz, and filter the EMG signal through band-pass filters with upper and lower cut-off frequencies of 3 Hz and 12 Hz to obtain the filtered attitude signal and muscle. Electric signal; set the current coordinate system as the earth coordinate system x, y, z, and calculate the attitude angle change according to the attitude signal. The specific method is as follows:

Define attitude angles θ ₁ , θ ₂ , θ ₃ , attitude angle changes Δθ ₁ , Δθ ₂ , Δθ ₃ , where θ ₁ , θ ₂ , θ ₃ are elbow flexion-extension angle, lower arm internal rotation, respectively -External rotation angle, wrist flexion-extension angle, △θ ₁ , △θ ₂ , △θ ₃ are the change in elbow joint flexion-extension angle, the lower arm inward-external rotation angle change, and the wrist joint flexion-extension angle change quantity;

According to the obtained filtered attitude signal, that is, the attitude quaternion of the upper arm, lower arm and hand IMU sensor, respectively calculate the projection α _x of the attitude angle of the upper arm, lower arm and hand IMU sensor on the coordinate system x, y and z axes, α _y , α _z , β _x , β _y , β _z , γ _x , γ _y , γ _z ;

From the arm geometry we get:

△θ _1x =△α _x +△β _x △θ _3x =△β _x +△γ _x

△θ _1y =△α _y +△β _y , △θ _3y =△β _y +△γ _y

△θ _1z =△α _z +△β _z △θ _3z =△β _z +△γ _z

Among them, △θ _1x , △θ _1y , △θ _1z , △θ _3x , △θ _3y , and △θ _3z are the elbow flexion-extension angle variation Δθ ₁ and the wrist flexion-extension angle variation Δθ ₃ respectively. The projection of the x, y, z axes of the coordinate system; △α _x , △α _y , △α _z is the angle change of the projection of the upper arm attitude angle on the x, y, z axes; △β _x , △β _y , △β _z is the angular change of the projection of the lower arm attitude angle on the x, y, z axis; △γ _x , △γ _y , △γ _z is the angular change of the projection of the hand attitude angle on the x, y, z axis;

Calculate the attitude angle variation Δθ ₁ , Δθ ₃ respectively, the expressions are as follows:

Since the internal rotation or external rotation of the lower arm is rotated along the axis of the elbow joint and the wrist joint, the attitude angle change _Δθ2 is directly derived from the quaternion output by the lower arm IMU sensor. According to the conversion relationship between the quaternion and the attitude angle, the output quaternion can be converted into the angle change.

中；按时间轴分别剔除数据集

和

的前后10％的数据，更新数据集

和

。Step 3, save the discrete data of the attitude angle change and the discrete data of the EMG signal output in the step 2 in the data set respectively.

Medium; cull datasets separately by time axis

and

10% of the data before and after, update the dataset

and

.

和

分别进行等步长窗口化处理，为了连续性描述震颤的变化，将数据集

和

的步长设置为100ms，窗口时长为2s，数据交叠率为95％；将处理后的数据集

和

组合，选取其中70％作为训练集，用于HMM模型参数学习，其余的30％作为测试集，用于HMM模型识别。Step 4, the dataset

and

The equal-step window processing is performed separately. In order to continuously describe the change of tremor, the data set is

and

The step size is set to 100ms, the window duration is 2s, and the data overlap rate is 95%; the processed data set

and

Combined, 70% of them are selected as the training set for HMM model parameter learning, and the remaining 30% are used as the test set for HMM model recognition.

Step 5: Extract the time-domain and frequency-domain characteristics of the attitude angle change of Parkinson's tremor in the resting state and the EMG signal characteristics from the windowed data in the training set and the test set respectively; the method is as follows:

Extract the time-domain and frequency-domain features of Parkinson's tremor in the resting state of attitude angle changes, as follows:

描述震颤程度，分别计算姿态角θ₁,θ₂,θ₃的平均角变化率

单位为度/秒；计算公式如下：Feature 1: Use the average angular change rate of Parkinson's tremor in the angular change time domain of the resting state

Describe the degree of tremor and calculate the average angular change rate of attitude angles θ ₁ , θ ₂ , θ ₃ respectively

The unit is degrees per second; the calculation formula is as follows:

where n represents the data length, n=t*f _s , t is the window duration, and f _s is the sampling frequency of the IMU sensor;

Feature 2: Use the energy of Parkinson's tremor in the resting state angle change frequency domain as a feature to describe the degree of tremor;

The attitude signal data in the training set is subjected to discrete Fourier transform, converted into frequency domain data, and the energy value is calculated:

where E is the energy value, f is the frequency point, P(f) is the signal energy spectral density, the unit is dB/sec; a, b are the frequency range boundary values of the resting tremor, respectively; [a, b] is the value of The typical frequency range of resting tremor is 3-6 Hz;

Feature 3: The information entropy of the angle change of Parkinson's tremor in the resting state is used as a feature to describe the degree of tremor; the information entropy T is defined as follows:

为频点f对应的出现概率，对数log取2为底。in

is the occurrence probability corresponding to the frequency point f, and the logarithm log takes 2 as the base.

Extracting EMG signal features to describe Parkinson's tremor, as follows:

Feature 4: Select the EMG kurtosis coefficient as a feature to describe Parkinson's tremor; define the kurtosis coefficient k as follows:

where μ is the data mean, Λ is the mathematical expectation, σ is the standard deviation, and x is the EMG discrete data;

即为过零率。所述归一化方法为：将肌电数据映射至[-1,1]区间，得到新的数据为：x′＝(x-x_mean)/(x_max-x_min)，其中x_mean为数据均值，x_max为最大值，x_min为最小值。Feature 5: Select the EMG signal zero-crossing rate as a feature to describe Parkinson's tremor; normalize the EMG signal standard, and measure the number of zeros in the normalized EMG data

is the zero-crossing rate. The normalization method is: map the EMG data to the [-1,1] interval, and obtain new data as: x′=(xx _mean )/(x _max -x _min ), where x _mean is the data mean , x _max is the maximum value, and x _min is the minimum value.

Step 6: Define the hidden state set of the HMM model according to the UPDRS Parkinson's rating scale; set the tremor amplitude to be h, and the hidden state set includes:

state 0: no tremor;

State 1: slight tremor, that is, tremor occurs and 0<h≤1cm;

State 2: Moderate tremor, that is, tremor occurs and 1<h≤3cm;

State 3: severe tremor, that is, tremor occurs and 3<h≤10cm;

State 4: Severe tremor, that is, tremor occurs and h>10cm;

where h is the tremor amplitude;

According to the features extracted from the training set, state labels 0, 1, 2, 3, and 4 are set for each data segment respectively.

Step 7: Initialize the HMM model parameters, the number of HMM hidden states M=5, the initial probability π=[1,0,0,0,0], the state observation probability B is represented by a set of mixed Gaussian densities; let the parameter λ=( A, B, π), A represents the hidden state transition probability; the feature parameters extracted from the training set are used as the training observation sequence, the Baum-Welch algorithm is used for model training, and the optimal parameter λ of the HMM model is iteratively calculated to obtain the trained HMM. Model; the features include angle change features and myoelectric signal features.

Step 8: Take the feature parameter Λ′ extracted from the test set as the test observation sequence, input it into the trained HMM model, use the Viterbi algorithm to solve the hidden state sequence when P(Λ′|λ) is the largest, and output the current optimal Hidden state (ie 0, 1, 2, 3, 4); among them, the feature parameter Λ' is the angle feature and the EMG feature set; P(Λ'|λ) means that when the Viterbi algorithm solves the given HMM model parameter λ, The probability of occurrence of the observation sequence Λ'.

According to research, most resting tremors are manifested in the upper limbs of the human body. Therefore, the present invention mainly selects the upper limbs for analysis. The human arm can be modeled as a kinematic chain, as shown in Figure 1, consisting of three arm segments (upper arm, lower arm, and hand) and three joints (shoulder, elbow, and wrist), and three IMUs. The sensors were deployed on the upper arm near the elbow (above the distal humerus, position 1 in the figure), the lower arm near the wrist (above the distal radius and ulna, position 3 in the figure), and the dorsal surface of the hand (position 4 in the figure). The soft tissue of the arm has less influence on these positions, and the signal is less susceptible to disturbance by soft tissue micromotion. At the same time, the EMG sensor measurement electrode is deployed on the surface of the upper arm biceps brachii (position 2 in the figure). This is because the elbow flexion-extension and internal rotation-external rotation have a direct relationship with the biceps brachii, so this position needs to be measured The surface EMG signal is used for the later EMG feature extraction. Because the resting state tremor usually manifests as the combined motion of the arm wrist joint and the elbow joint rotation, the present invention selects three angles describing the rotation feature as the input quantities of the later feature extraction, which are respectively θ ₁ (the elbow joint bending-extension angle, Angle 5 in the figure), θ ₂ (inward-outer rotation angle of the lower arm, angle 6 in the figure) and θ ₃ (hand flexion-extension angle, angle 7 in the figure).

FIG. 2 shows the overall flow chart of the tremor evaluation method in the present invention, which is mainly divided into three parts, which are data collection and preprocessing, feature extraction, and HMM model training and recognition. The calculation process of the model is divided into an offline stage and an online stage. The offline stage is used for tremor data segmentation and tremor feature dataset construction; the online stage is used for tremor feature identification and state evaluation. Compared with the conventional threshold evaluation method, the HMM method has the following advantages: the Markov chain structure can retain the structural information of the tremor feature sequence; the HMM model parameters can represent the statistical characteristics of the tremor; as a probabilistic model, the HMM is no longer required. Set the threshold. Therefore, the present invention can directly reflect the transition process of the tremor state from the time sequence by using the HMM model, and can better describe the time sequence characteristics of the tremor of Parkinson's patients in daily life.

Compared with other types of time series signal analysis models, Hidden Markov Model (HMM) can better describe the characteristics of time-related signals and can be well characterized as a random process of parameters. It preserves the structural information of the feature signal without the need for thresholds used in heuristic rules. In order to better describe the tremor state transition and improve the generalization ability of the model, the present invention selects the angular feature information and the EMG feature information as the model input, the feature matrix calculation amount is small, and it is more suitable for the wearable somatosensory network. Occasions with high energy efficiency requirements.

According to the actual process of the tremor state transition, the present invention only selects the continuous and progressive state transition process, as shown in Figure 3, namely 0-0, 0-1, 1-1, 1-2, 2-2, 2-1, 2 -3, 3-3, 3-2, 3-4, 4-4, 4-3, discard the tremor state jump process 0-2, 2-0, 0-4, 1-3, 4-2, etc. ; In addition, it is defined that the initial state and the end state of the dataset are both 0 states, that is, starting from the tremor-free state at first, until the end of the tremor-free state.

Figure 4 shows the wrist angle change amount, the average angle change rate, the angle frequency domain energy value, the angle frequency domain information entropy and the hidden state sequence identified by the HMM model from top to bottom. According to the actual measurement results, it can be concluded that the optimal output sequence of the HMM model is: 0-1-2-1-0. It can be seen that the angle feature selected by the present invention has good descriptiveness and discrimination, and has a good recognition effect for Parkinson's tremor evaluation. Through the method of the present invention, the state recognition rate can be achieved over 98%.

Claims (10)

1. a Parkinson's resting state tremor assessment method based on wearable somatosensory net, is characterized in that: the method comprises the following steps: Step 1, deploy sensors on human hands and arms, initialize the sensors, eliminate sensor drift, perform initial calibration, and set sensor sampling frequencies; collect human body posture signals and human body surface EMG signals through the deployed sensors; Step 2: Filter the attitude signal and the myoelectric signal through a band-pass filter according to the frequency response characteristic of the tremor in the resting state, to obtain the filtered attitude signal and the myoelectric signal, and calculate the attitude angle change according to the attitude signal; Step 3, save the discrete data of attitude angle change and the discrete data of EMG signal output in step 2 in data sets N and C respectively; remove the data before and after p% of data sets N and C respectively according to the time axis, and update the data set N and C; Step 4. Perform equal-step windowing processing on datasets N and C respectively, and combine the processed datasets N and C, and select q% of them as the training set for HMM model parameter learning, and the rest 1-q % as a test set for HMM model recognition; Step 5: Extract the time-domain and frequency-domain characteristics of the attitude angle change of Parkinson's tremor in the resting state and the EMG signal characteristics from the windowed data in the training set and the test set respectively, including the average angular change rate and angle in the time domain of the angle change. Energy in changing frequency domain, information entropy of angle change, EMG signal kurtosis coefficient, EMG signal zero-crossing rate; Step 6, according to the UPDRS Parkinson's assessment scale, define the hidden state set of the HMM model; according to the features extracted from the training set, set state labels for each data segment respectively; Step 7, initialize the HMM model parameters, use the feature parameters extracted from the training set as the training observation sequence, carry out model training, calculate the optimal parameters of the HMM model, and obtain a trained HMM model; Step 8, take the feature parameters extracted from the test set as the test observation sequence, input it into the trained HMM model, solve the hidden state sequence when the probability of occurrence of the observation sequence is the largest, output the current optimal hidden state, and obtain the tremor state evaluation result. 2. a kind of Parkinson's resting state tremor assessment method based on wearable somatosensory net according to claim 1, is characterized in that: described step 1, respectively deploy IMU sensor on upper arm, lower arm, hand, simultaneously The EMG sensor is deployed on the upper arm; the human body attitude data is collected by the IMU sensor, and the electrical signal of the human body surface is collected by the EMG sensor; the IMU sensor includes a three-axis accelerometer, a three-axis gyroscope and a three-axis magnetometer, and the attitude signal collected by the IMU sensor Output as a quaternion. 3. A method for evaluating Parkinson's resting state tremor based on a wearable somatosensory network according to claim 2, wherein the sampling frequency of the IMU sensor is set to 100 Hz, and the sampling frequency of the EMG sensor is 50 Hz. 4. a kind of Parkinson's resting state tremor assessment method based on wearable somatosensory net according to claim 2, it is characterized in that: select the nine-axis IMU sensor whose model is GY953, and the EMG sensor model is OYmotion. 5. a kind of Parkinson's resting state tremor assessment method based on wearable somatosensory net according to claim 2, is characterized in that: described step 2, set current coordinate system to be the earth coordinate system x, y, z, Calculate the attitude angle change as follows: Define attitude angles θ ₁ , θ ₂ , θ ₃ , attitude angle changes Δθ ₁ , Δθ ₂ , Δθ ₃ , where θ ₁ , θ ₂ , θ ₃ are elbow flexion-extension angle, lower arm internal rotation-external rotation angle, respectively , wrist flexion-extension angle, Δθ ₁ , Δθ ₂ , Δθ ₃ are the change of elbow joint flexion-extension angle, the change of lower arm internal rotation-external rotation angle, and the change of wrist joint flexion-extension angle; According to the attitude quaternion of the upper arm, lower arm and hand IMU sensor, respectively calculate the projection α _x , α _y , α _z , β _x of the attitude angle of the upper arm, lower arm and hand IMU sensor on the x, y, z axes of the coordinate system, β _y , β _z , γ _x , γ _y , γ _z ; From the arm geometry we get: Δθ _1x =Δα _x +Δβ _x Δθ _3x =Δβ _x +Δγ _x Δθ _1y =Δα _y +Δβ _y , Δθ _3y =Δβ _y +Δγ _y Δθ _1z =Δα _z +Δβ _z Δθ _3z =Δβ _z +Δγ _z Among them Δθ _1x , Δθ _1y , Δθ _1z , Δθ _3x , Δθ _3y , Δθ _3z are the elbow flexion-extension angle change Δθ ₁ , the wrist flexion-extension angle change Δθ ₃ in the coordinate system x, y, z axes, respectively Δα _x , Δα _y , Δα _z are the angular changes of the projection of the upper arm attitude angle on the x, y, z axes; Δβ _x , Δβ _y , Δβ _z are the lower arm attitude angles on the x, y, z axes The angle change of the projection; Δγ _x , Δγ _y , Δγ _z is the angle change of the projection of the hand attitude angle on the x, y, and z axes; Calculate the attitude angle changes Δθ ₁ , Δθ ₃ respectively, the expressions are as follows:

The attitude angle change Δθ ₂ is directly derived from the quaternion output by the lower arm IMU sensor. 6. a kind of Parkinson's resting state tremor assessment method based on wearable somatosensory net according to claim 5, is characterized in that: described step 5, extracts Parkinson's tremor in the resting state attitude angle change time domain and frequency domain features, as follows: Feature 1: Use the average angular change rate of Parkinson's tremor in the angular change time domain of the resting state

Describe the degree of tremor and calculate the average angular change rate of attitude angles θ ₁ , θ ₂ , θ ₃ respectively

Calculated as follows:

where n represents the data length, n=t*f _s , t is the window duration, and f _s is the sampling frequency of the IMU sensor; Feature 2: Use the energy of Parkinson's tremor in the resting state angle change frequency domain as a feature to describe the degree of tremor; The attitude signal data in the training set is subjected to discrete Fourier transform, converted into frequency domain data, and the energy value is calculated:

where E is the energy value, f is the frequency point, P(f) is the signal energy spectral density; a and b are the boundary values of the frequency range of the resting tremor, respectively; Feature 3: The information entropy of the angle change of Parkinson's tremor in the resting state is used as a feature to describe the degree of tremor; the information entropy T is defined as follows:

in

is the occurrence probability corresponding to the frequency point f, and the logarithm log takes 2 as the base. 7. a kind of Parkinson's resting state tremor assessment method based on wearable somatosensory net according to claim 1, it is characterized in that: described step 5, extracts myoelectric signal feature description Parkinson's tremor, is specifically as follows: Feature 4: Select the EMG kurtosis coefficient as a feature to describe Parkinson's tremor; define the kurtosis coefficient k as follows:

where μ is the data mean, Λ is the mathematical expectation, σ is the standard deviation, and x is the EMG discrete data; Feature 5: Select the EMG signal zero-crossing rate as a feature to describe Parkinson's tremor; normalize the EMG signal standard, and measure the number of zeros in the normalized EMG data

is the zero-crossing rate; the normalization method is: map the EMG data to the [-1,1] interval to obtain new data, and the calculation formula is: x′=(xx _mean )/(x _max -x _min ), where x _mean is the data mean, x _max is the maximum value, and x _min is the minimum value. 8. a kind of Parkinson's resting state tremor assessment method based on wearable somatosensory net according to claim 1, is characterized in that: described step 6, according to UPDRS Parkinson's assessment scale, define HMM model hidden state Set, let the tremor amplitude be h, the hidden state set includes: state 0: no tremor; State 1: slight tremor, that is, tremor occurs and 0<h≤1cm; State 2: Moderate tremor, that is, tremor occurs and 1<h≤3cm; State 3: severe tremor, that is, tremor occurs and 3<h≤10cm; State 4: Severe tremor, that is, tremor occurs and h>10cm; According to the features extracted from the training set, state labels 0, 1, 2, 3, and 4 are set for each data segment respectively. 9. A wearable somatosensory net-based Parkinson's resting state tremor assessment method according to claim 1, wherein in step 7, the HMM model parameters include HMM hidden state number M, initial Probability π, state observation probability B; initialize HMM implicit state number M=5, initial probability π=[1,0,0,0,0], state observation probability B is represented by a set of mixed Gaussian densities; let the parameter λ= (A, B, π), A represents the implicit state transition probability; The feature parameters extracted from the training set are used as the training observation sequence, the Baum-Welch algorithm is used for model training, and the optimal parameter λ of the HMM model is iteratively calculated. 10. A wearable somatosensory net-based Parkinson's resting state tremor assessment method according to any one of claims 1-9, wherein in the step 8, the feature parameter Λ' extracted in the test set is used as Test the observation sequence, input the trained HMM model, use the Viterbi algorithm to solve the hidden state sequence when P(Λ′|λ) is the largest, and output the current optimal hidden state; among them, the feature parameter Λ′ is the angle feature and EMG feature set; P(Λ′|λ) represents the probability of the occurrence of the observation sequence Λ′ when the Viterbi algorithm solves the given HMM model parameter λ.

Images