CN114869272A - Postural tremor detection model, posture tremor detection algorithm, and posture tremor detection device - Google Patents

Abstract

The present application discloses a posture tremor detection model, which is based on an ensemble learning model, and the input features of the model are multiple; at least one input feature of the multiple input features is obtained from the three-axis of the subject's hand acceleration data, at least one input feature is obtained from three-axis gyroscope data of the subject's hand, and at least one input feature is obtained from the three-axis magnetometer data of the subject's hand; the model is based on the multiple The input feature classifies the tremor level of the subject in a predetermined posture; the predetermined posture includes wing-beat position, arm extension position, and finger-pointing position. The present application utilizes various kinematic characteristics of accelerometer, gyroscope and magnetometer signals to effectively reveal the regularity and synchronization of tremor in patients with essential tremor, and can effectively characterize the characteristics of patients' tremor amplitude, tremor frequency, etc. A model for accurate quantitative grading of patients' tremor symptoms.

Description

Postural tremor detection model, postural tremor detection algorithm, and postural tremor detection equipment

technical field

The present application relates to pattern recognition, and in particular to a postural tremor detection algorithm based on wearable devices.

Background technique

Essential tremor (ET) is a degenerative neurological disease with characteristic motor symptoms. It is an autosomal dominant genetic disease, also known as familial or benign essential tremor. The most common tremor symptoms. About 60% of patients have a family history. Essential tremor is a monosymptomatic disease, and postural tremor is the only clinical manifestation of this disease. The so-called postural tremor refers to the tremor caused when the limb maintains a certain posture, and the tremor disappears naturally when the limb is completely relaxed. The tremor of this disease is common in the hands, followed by head tremor, and rarely in patients with lower extremity tremor.

The current clinical assessment is mainly based on expert consultation, combined with the review of the patient's chief complaint, and relies heavily on the professional knowledge and diagnostic experience of the physician. Research on objective quantification through wearable sensor technology combined with machine learning methods has great potential for application.

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.

Currently, medical-grade wearable systems for ET tremor severity, such as the Kinesia HomeView ^™ , require the patient to repeat a standardized tremor assessment maneuver on an hourly basis using a tablet computer as prompted. Alternatively, using an empirical sampling method application called PsyMate ^™ , users are required to fill out a questionnaire. Instead, these operations increase the burden on patients and interfere with daily life. In addition, other types of tremors and motor disturbances are easily mixed due to free movement in related studies.

In addition, the existing technical solution only uses the accelerometer to directly collect the tremor signal. The drawback of this solution is that the motion component is easily mixed into the activity, and the accumulated error of the sensor and the environmental noise also make the measured signal 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 signal collected by the accelerometer, and it is easy to be mistaken for a noise signal and discarded.

SUMMARY OF THE INVENTION

In view of the above problems, the present application aims to propose a postural tremor detection model, a postural tremor detection algorithm, and a postural tremor detection device, which perform quantitative assessment of tremor symptoms based on multiple signals.

The posture tremor detection model of the present application is based on an ensemble learning model,

There are multiple input features of the model; at least one input feature in the multiple input features is obtained from the triaxial acceleration data of the subject's hand, and at least one input feature is obtained from the three-axis acceleration data of the subject's hand. Axis gyroscope data, at least one input feature is obtained from triaxial magnetometer data of the subject's hand;

The model classifies the tremor level of the subject in a predetermined posture according to the plurality of input features;

The predetermined postures include a wing position, an arm extension position, and a pointing position.

Preferably, when detecting the wing position, the integrated learning model adopts the RUSBoost model based on the embedding method;

When the arm extension position is detected, the integrated learning model adopts the AdaBoost model based on the Wrapper method;

When detecting the finger position, the ensemble learning model adopts the Bagging model based on the mutual information method.

Preferably, the plurality of input features are selected from:

ACC_Max, ACC_Min, ACC_Mean, ACC_PeakP, ACC_Arv, ACC_Var, ACC_Std, ACC_Kurt, ACC_Skew, ACC_RMS, ACC_SF, ACC_CF, ACC_IF, ACC_LF, ACC_PwrP, ACC_PwrP_R, ACC_PrinP, ACC_SampEn, ACC_ApEn, ACC_FuzEn;

GYRO_Max, GYRO_Min, GYRO_Mean, GYRO_PeakP, GYRO_Arv, GYRO_Var, GYRO_Std, GYRO_Kurt, GYRO_Skew, GYRO_RMS, GYRO_SF, GYRO_CF, GYRO_IF, GYRO_LF, GYRO_PwrP, GYRO_PwrP_R, GYRO_PrinP, GYROG_SEnamp;

MAG_Max, MAG_Min, MAG_Mean, MAG_PeakP, MAG_Arv, MAG_Var, MAG_Std, MAG_Kurt, MAG_Skew, MAG_RMS, MAG_SF, MAG_CF, MAG_IF, MAG_LF, MAG_PwrP, MAG_PwrP_R, MAG_PrinP, MAG_SampEn, MAG_ApEn, MAG_FuzEn;

Among them, ACC_Max, ACC_Min, ACC_Mean, ACC_PeakP, ACC_Arv, ACC_Var, ACC_Std, ACC_Kurt, ACC_Skew, ACC_RMS, ACC_SF, ACC_CF, ACC_IF, ACC_LF, GYRO_Max, GYRO_Min, GYRO_Mean, GYRO_PeakP, GYRO_Arv, GYRO_Var, GYRO_Std, GYRO_Kurt, GYRO_Skew GYRO_SF, GYRO_CF, GYRO_IF, GYRO_LF, MAG_Max, MAG_Min, MAG_Mean, MAG_PeakP, MAG_Arv, MAG_Var, MAG_Std, MAG_Kurt, MAG_Skew, MAG_RMS, MAG_SF, MAG_CF, MAG_IF, MAG_LF are time domain features;

ACC_Max is the maximum signal peak value of the triaxial acceleration data; ACC_Min is the signal minimum peak value of the triaxial acceleration data; ACC_Mean is the signal average value of the triaxial acceleration data; ACC_PeakP is the signal peak-to-peak amplitude of the triaxial acceleration data; ACC_Arv is the signal peak value of the triaxial acceleration data. The average rectification value of the signal of the three-axis acceleration data; ACC_Var is the signal variance of the three-axis acceleration data; ACC_Std is the signal standard deviation of the three-axis acceleration data; ACC_Kurt is the signal kurtosis of the three-axis acceleration data; ACC_Skew is the signal kurtosis of the three-axis acceleration data degrees; ACC_RMS is the signal root mean square of the triaxial acceleration data; ACC_SF is the signal shape factor of the triaxial acceleration data; ACC_CF is the signal crest factor of the triaxial acceleration data; ACC_IF is the signal pulse factor of the triaxial acceleration data; ACC_LF is the signal pulse factor of the triaxial acceleration data Signal margin factor for axis acceleration data;

GYRO_Max is the maximum peak value of the triaxial gyroscope data; GYRO_Min is the minimum peak value of the triaxial gyroscope data; GYRO_Mean is the signal average value of the triaxial gyroscope data; GYRO_PeakP is the signal peak-to-peak amplitude of the triaxial gyroscope data; GYRO_Arv is the average rectification value of the triaxial gyroscope data; GYRO_Var is the signal variance of the triaxial gyroscope data; GYRO_Std is the signal standard deviation of the triaxial gyroscope data; GYRO_Kurt is the signal kurtosis of the triaxial gyroscope data; GYRO_Skew is the signal kurtosis of the triaxial gyroscope data The signal kurtosis of the triaxial gyroscope data; GYRO_RMS is the signal root mean square of the triaxial gyroscope data; GYRO_SF is the signal form factor of the triaxial gyroscope data; GYRO_CF is the signal crest factor of the triaxial gyroscope data; GYRO_IF is the signal crest factor of the triaxial gyroscope data; Signal pulse factor of axis gyroscope data; GYRO_LF is the signal margin factor of three-axis gyroscope data;

MAG_Max is the maximum signal peak value of the triaxial magnetometer data; MAG_Min is the minimum peak value of the triaxial magnetometer data; MAG_Mean is the signal average value of the triaxial magnetometer data; MAG_PeakP is the signal peak-to-peak amplitude of the triaxial magnetometer data; MAG_Arv is the signal average rectification value of the triaxial magnetometer data; MAG_Var is the signal variance of the triaxial magnetometer data; MAG_Std is the signal standard deviation of the triaxial magnetometer data; MAG_Kurt is the signal kurtosis of the triaxial magnetometer data; MAG_Skew is the The signal kurtosis of the triaxial magnetometer data; MAG_RMS is the signal root mean square of the triaxial magnetometer data; MAG_SF is the signal form factor of the triaxial magnetometer data; MAG_CF is the signal crest factor of the triaxial magnetometer data; MAG_IF is the signal crest factor of the triaxial magnetometer data Signal pulse factor of axis magnetometer data; MAG_LF is the signal margin factor of triaxial magnetometer data;

ACC_PwrP, ACC_PwrP_R, ACC_PrinP, GYRO_PwrP, GYRO_PwrP_R, GYRO_PrinP, MAG_PwrP, MAG_PwrP_R, MAG_PrinP are frequency domain features;

ACC_PwrP is the peak power of the triaxial acceleration data; ACC_PwrP_R is the peak power ratio of the triaxial acceleration data; ACC_PrinP is the main peak power of the triaxial acceleration data;

GYRO_PwrP is the peak power of the three-axis gyroscope data; GYRO_PwrP_R is the peak power ratio of the three-axis gyroscope data; GYRO_PrinP is the main peak power of the three-axis gyroscope data;

MAG_PwrP is the peak power of the triaxial magnetometer data; MAG_PwrP_R is the peak power ratio of the triaxial magnetometer data; MAG_PrinP is the main peak power of the triaxial magnetometer data;

ACC_SampEn, ACC_ApEn, ACC_FuzEn, GYRO_SampEn, GYRO_ApEn, GYRO_FuzEn, MAG_SampEn, MAG_ApEn, MAG_FuzEn are nonlinear features;

ACC_SampEn is the sample entropy of the triaxial acceleration data; ACC_ApEn is the approximate entropy of the triaxial acceleration data; ACC_FuzEn is the fuzzy entropy of the triaxial acceleration data;

GYRO_SampEn is the sample entropy of the three-axis gyro data, GYRO_ApEn is the approximate entropy of the three-axis gyro data; GYRO_FuzEn is the fuzzy entropy of the three-axis gyro data.

Preferably, the maximum number of splits of the RUSBoost model based on the embedding method is 311, the number of learners is 483, and the learning rate is 0.8578; the multiple input features are: ACC_RMS; ACC_Min; ACC_Arv; ACC_LF; GYRO_RMS; GYRO_Max; GYRO_Min; GYRO_Arv; GYRO_SF; GYRO_LF; MAG_FuzEn; MAG_RMS; MAG_Max; MAG_Mean; MAG_SF; MAG_CF.

Preferably, the maximum number of splits of the AdaBoost model based on the Wrapper method is 212, the number of learners is 45, and the learning rate is 0.9986; the multiple input features are: ACC_RMS; ACC_Min; ACC_Arv; ACC_LF; GYRO_Max; GYRO_Min; GYRO_Arv; GYRO_LF; GYRO_RMS; MAG_FuzEn; MAG_RMS; MAG_Max.

Preferably, the maximum number of splits of the bagging model based on the mutual information method is 174, the number of learners is 362, the learning rate is 0.8578, and 4 input features are randomly selected from the plurality of input features; the plurality of input features ACC_Max; ACC_Mean; ACC_SF; ACC_CF; ACC_IF; ACC_LF; GYRO_Max; GYRO_Min; GYRO_Mean; GYRO_PeakP;

A posture tremor detection algorithm of the present application obtains the triaxial acceleration data, triaxial acceleration data, triaxial acceleration data, triaxial acceleration data, triaxial acceleration data, triaxial acceleration data, triaxial acceleration data, triaxial acceleration data, triaxial acceleration data, triaxial acceleration data, triaxial acceleration data, triaxial acceleration data, triaxial acceleration data, triaxial acceleration data, triaxial Gyroscope data, triaxial magnetometer data; after processing triaxial acceleration data, triaxial gyroscope data, triaxial magnetometer data to extract multiple input features; using the multiple input features as the utilization claim 1- The input feature of any one of the postural tremor detection models in 6, to classify the tremor level of the subject.

The posture tremor detection device of the present application includes a computing unit;

The computing unit runs the postural tremor detection model of any one of claims 1-6 to classify the tremor level of the subject.

Preferably, it includes a three-axis accelerometer, a three-axis gyroscope, and a three-axis magnetometer; the three-axis accelerometer, three-axis gyroscope, and three-axis magnetometer are arranged on the subject's hand; The three-axis acceleration data is obtained through the three-axis accelerometer; the three-axis gyroscope data is obtained through the three-axis gyroscope; and the three-axis magnetometer data is obtained through the three-axis magnetometer.

Preferably, the three-axis accelerometer, the three-axis gyroscope, and the three-axis magnetometer are wearable devices, and the three-axis acceleration data, the three-axis gyroscope data, the three-axis The magnetometer data is sent to the computing unit.

This application utilizes various kinematic characteristics of accelerometer, gyroscope and magnetometer signals such as root mean square (RMS), kurtosis, form factor, spectral peak, peak power, peak ratio, approximate entropy, sample entropy and blur Entropy and other characteristics), effectively revealing the regularity and synchronization of tremor in patients with essential tremor, and can effectively characterize the characteristics of patients' tremor amplitude, tremor frequency, etc., and establish a model that can accurately quantify and classify patients' tremor symptoms.

Description of drawings

Figure 1 is a structure and experimental demonstration of the wearable device used in this application.

FIG. 2 is the overall experimental flow of the present invention.

Figure 3 shows the power spectral density function calculated from the vector amplitude of the acceleration signal (the intercepted 15s signal segment).

Figure 4 shows the ROC curves of the five classification models.

Figure 5 shows the confusion matrix of the best five-class model.

Detailed ways

Hereinafter, the present application will be described in detail with reference to the accompanying drawings.

In Figure 1, (a) shows the basic structure of an IMU-based wearable device. (b) shows the acquisition process of the experiment. (c)-(e) show three postural tremor tasks. In Figure 2, an automatic scoring system for orthostatic tremor in ET patients is constructed based on data processing and data analysis. In Figure 3, a method based on the power spectrum tolerance span is used to find the peak power, and the shaded area shows the peak power of the tremor signal. In Fig. 4, (a-c) ROC curves of five machine learning models on all feature sets, respectively, in the wing stroke, arm extension, and finger-to-finger tasks. (d-f) ROC curves of comprehensive learning tree models based on five feature selection methods in three tasks. In Fig. 5, (a-c) represent the confusion matrices of the optimal classifier AdaBoost overall feature set in the wing beat, arm extension, and finger-to-finger tasks. (d-f) represents the confusion matrix of the optimal ensemble learning tree model based on the optimal feature selection method in the three tasks. Among them, (d) represents the classification result of RUSBoost based on the Embedded method, and the specific parameters are: the maximum number of splits is 311, the number of learners is 483, and the learning rate is 0.8578; (e) represents the classification result of AdaBoost based on the Wrapper method, and the specific parameters is: the maximum number of splits is 212, the number of learners is 45, and the learning rate is 0.9986; (f) represents the classification result of bagging based on the mutual information method. The specific parameters are that the maximum number of splits is 174, the number of learners is 362, and the learning rate is 0.8578 , the training basic estimator needs to randomly select 4 features.

The specific implementation steps of the method of the present invention are:

(1) Deploy the sensor on the human hand, initialize the sensor through the host computer (or computing unit), eliminate sensor drift, perform initial calibration, and set the sampling frequency of the sensor; collect the posture signal of the human hand through the deployed sensor. Specifically, the wearable device based on the nine-axis IMU is used to collect the hand tremor data of the patient's hands under the specified action; at the same time, the CRST scale is used to score the completion of the patient's specified action; A neurologist was blindly reviewed and scored, and a professional neurologist videotaped the patient's movements during the laboratory examination.

Laboratory examinations were conducted by a neurologist with expertise in movement disorders, and video data were recorded (CMOS camera, 48MP, 1920*1080 HD, 60 frames/sec) to support independent scoring by three neurologists (Blind points to each other). During laboratory tests, patients wear tiny posture sensors on the back of their hands, and digital signals are transmitted wirelessly in a host computer via a Bluetooth connection. The host computer displays the signal waveform change in real time and stores it on the computer hard disk in the form of text. In order to ensure that valid chatter signals are resolved, the sampling frequency of the device is set to 100Hz and the baud rate is set to 115,200Bytes per second.

Scores are jointly scored by two experts. If the scores are inconsistent, a third expert needs to be asked to make a final judgment on the scores by viewing the video. The specific scoring process is as follows. 1) Two neurologists rated the severity of postural tremor in patients by watching video materials, and obtained two mutually blinded score sheets. 2) The data analysis engineer will count the consistent scores of the two experts. For actions with inconsistent scores, the video data will be finalized by another experienced neurologist until a reliable score is obtained. This experimental design also avoids training errors due to systemic bias, making machine learning models more reliable.

The IMU includes a device box, a nine-axis inertial sensor, an embedded wireless module, a lithium battery, a power button, a status indicator, and a data cable. The embedded wireless module, lithium battery, power button, and status indicator are placed in the device box. The power button interface and status indicator interface are left on the surface of the device box, and the data cable connection interface is left on the side of the device box. The nine-axis inertial sensor and the device The boxes are connected by a data cable; the lithium battery is responsible for the power supply of the embedded wireless module and the status indicator; the power button controls the on and off of the lithium battery supply; the nine-axis inertial The tremor data of the hand under the action, and the tremor data is transmitted through the embedded wireless module and transmitted to the host computer (or computing unit); the tremor data refers to the three-axis acceleration data obtained by the nine-axis inertial Axis gyroscope data and three-axis magnetometer data.

The core of the data acquisition of the wearable system used in the present invention is the inertial sensing unit. A QFN packaged multi-chip (MCM) MPU-9250 (InvenSense, USA) that integrates a three-axis accelerometer, a three-axis gyroscope, and a three-axis magnetometer AK8963 (AKM Semiconductor, USA) via an embedded microcontroller (MSP430). The monitored digital signal is wirelessly transmitted through the Bluetooth protocol (BR-LE4.0-D2). The capacitance on each axis of the accelerator measures the deviation of each axis; the gyroscope measures the angular velocity of each axis based on the Coriolis effect; the magnetometer collects the electromagnetic intensity of the earth's magnetic field based on the Hall effect, and can estimate the carrier's angular velocity. attitude, orientation, etc. In addition, the host of the data acquisition system can be used to set the acquisition frequency, calibrate the sensor, visualize the acquired signal, and store the data (Fig. 1(a)).

The specified action refers to a postural tremor action, that is, a tremor symptom that occurs when the affected part of the body actively maintains a specific posture. According to Part A of the CRST Standard Scale (Figure 1(b-e)), patients were asked to sit comfortably in a chair with eyes looking forward, arms outstretched, wrists slightly extended, fingers apart, and then maintain three positions in sequence. 1) Wing stroke position, arms outstretched, arms slightly extended, fingers separated; 2) Arm extension position, arms straight, palms down; 3) Finger position, arms outstretched, arms straight, palms down down. The index fingers of the left and right hands face each other, palms facing inward.

Using the CRST scale to score the completion of the patient's specified action means that the professional physician judges the patient's tremor level of resting tremor according to the completion of the patient's specified action. , Level 2, Level 3, Level 4. The CRST scale shows that: Postural tremor is classified into 5 grades of tremor symptoms: 1) None; 2) Mild, sometimes occurring; 3) Moderate amplitude, occurring during movement; 4) Moderate amplitude, occurring during fixed movements; 5) The magnitude is large and affects eating.

(2) Figure 2 shows the flow of data analysis. It is mainly divided into two parts: data processing and feature extraction. In the method of the present invention, the collected raw IMU data is filtered and sliced, and converted into a clean isometric sequence for feature extraction. In the laboratory examination of patients with essential tremor, even if the examination actions are strictly regulated in accordance with the requirements of professional doctors, the human body will produce conscious movements, which are usually manifested as low-frequency components of the signal, while high-frequency components include inter-motion tremor and noise. Specifically, the attitude sensing signals obtained experimentally include ET tremor, physiological tremor and various random noises. Before feature extraction, it is necessary to filter out interference signals other than essential tremor as much as possible, such as subtle physiological tremor caused by normal muscle activation. In contrast, ET tremors will be much larger in amplitude, with a frequency of 4-12 Hz. Therefore, we ensure stable sensor signal output in dynamic environment by integrating the dynamic Kalman filtering algorithm on the hardware side. In addition, the attitude solver can filter the random errors of the sensing element (including zero point deviation, temperature drift, inter-axis alignment error, etc.) at the front end to ensure the reliability of the signal and preserve the high signal-to-noise ratio of the oscillatory signal. Although stable postural signals can be obtained, small-amplitude human intentional movements and physiologic tremors are also mixed during tremor movements. Wavelet transform is suitable for non-stationary signals, and has good time-frequency localization characteristics in the problems of signal mutation, compression reconstruction and denoising. Therefore, the method of the present invention selects the sym3 wavelet, performs two-level independent decomposition of the noisy tremor signal based on the soft threshold function, and reconstructs useful pathological tremor information.

(3) Since the early and late stages of acquisition are easily affected by test preparation and test stop state transitions, the data of C% before and after the filtered data are removed according to the time axis, and the data of C% to 1-C% according to the center of the time axis are retained. Therefore, the stable signal component can be intercepted from the filtered data including the attitude component and the tremor component output in step (2); preferably, the value of C% is 5%.

(4) Based on the tremor data after filtering, perform equal-length sliding window data amplification to keep the tremor data at the same length, and then calculate the time-domain amplitude change, frequency-domain peak power change and nonlinear entropy value change of the tremor data. and other characteristics; preferably, in order to describe the change of tremor continuously, the moving step of the sliding window is set to 1s, the window duration is 4s, and the data overlap rate is 75%.

(5) Randomly construct a sample training set and a sample test set in an appropriate proportion for the filter-processed feature sets of all patients' designated action tremors; use the expert scores obtained in step (1) to set the tremor severity for each segment of data Tag of.

In order to ensure the generalization of the training model, a five-fold cross-validation with a data split ratio of 4:1 was selected.

(6) Build five classification models of various machine learning algorithms, including support vector machines, ensemble tree models, linear discriminant analysis, naive Bayesian models, and K-nearest neighbor algorithms, and use optimization methods to optimize and adjust parameters to obtain the best classification 's model.

Select the Bayesian optimization algorithm to find the optimal solution with probability, then use grid search to adjust parameters in a small range, and select five-fold cross-validation in the training phase to improve the generalization ability of the model.

(7) Validation of the method, including: verifying the relationship between the seven parameters obtained in (2) and (4) above and the tremor level and verifying the accuracy of the classifier constructed based on the machine learning model. The performance evaluation index gives the optimal classification model.

A similar database can be established to verify the performance of the model proposed by the method of the present invention. Preferably, it can be set according to the parameters specified in the steps of the present invention. It is recommended to set up an age-matched gender-balanced database of no less than 100 people (average age is about 60 years old, half of them are male), the average recording time is 6 minutes, and the sensor sampling frequency is set to 100Hz.

The overall flow of the implementation of the present invention is shown in FIG. 2 .

In step (4), feature preprocessing is first performed: in order to achieve simple and effective state recognition and reduce the dependence on the wearing position and action process, the signal vector magnitude (SVM) is calculated separately for the three-axis sequence of each sensor. SVM can also reduce the complexity of vector operations for each IMU signal-sensitive axis from a macro perspective. Specifically, three-axis acceleration data is used to illustrate the data analysis and feature extraction process. a _svm (i) represents the collective acceleration of the ith sampling point, and its calculation formula is:

(1)

where a _x (i), a _y (i), and a _z (i) are the accelerations on the x, y, and z axes of the ith sampling point, respectively. For a sampled data segment of length N, the final SVM acceleration sequence is a _svm ={a _svm (0),a _svm (1),...,a _svm (N-1)}.

Further, using the SVM sequence obtained after feature preprocessing, the time domain, frequency domain and nonlinear features of the posture tremor state of the resting tremor are extracted from the accelerometer signal, gyroscope signal, and magnetometer signal respectively to describe the tremor severity. details as follows:

Time-domain features: The amplitude of tremor in motion disorders is relatively significant at rest. The method of the present invention extracts 20 interpretable features from the SVM sequence of three-channel sensing signals (acceleration, angular velocity and angle), including Time domain, frequency domain and nonlinear features. The specific description of the characteristic parameters is shown in Table 1. The time domain parameters are all morphological characteristics of the signal. Among them, the skewness and the kurtosis are the third-order normalized moment and the fourth-order normalized moment of the signal, respectively, and the specific formula is shown in formula (2-3).

Among them, a _mean represents the sample mean of the acceleration SVM sequence a _svm .

Table 1 The characteristic matrix and its definition of the sensing signal of each channel in the experimental stage

In summary, according to Table 1, the formulas in Table 1 are applied to the three-axis acceleration signal, the three-axis gyroscope signal, and the three-axis magnetometer signal, and 60 input features can be obtained, which are listed as follows:

ACC_Max,ACC_Min,ACC_Mean,ACC_PeakP,ACC_Arv,ACC_Var,ACC_Std,ACC_Kurt,ACC_Skew,ACC_RMS,ACC_SF,ACC_CF,ACC_IF,ACC_LF,ACC_PwrP,ACC_PwrP_R,ACC_PrinP,ACC_SampEn,ACC_ApEn,ACC_FuzEn;

GYRO_Max,GYRO_Min,GYRO_Mean,GYRO_PeakP,GYRO_Arv,GYRO_Var,GYRO_Std,GYRO_Kurt,GYRO_Skew,GYRO_RMS,GYRO_SF,GYRO_CF,GYRO_IF,GYRO_LF,GYRO_PwrP,GYRO_PwrP_R,GYRO_PrinP,GYROG_En,ROG_Samp;;

MAG_Max, MAG_Min, MAG_Mean, MAG_PeakP, MAG_Arv, MAG_Var, MAG_Std, MAG_Kurt, MAG_Skew, MAG_RMS, MAG_SF, MAG_CF, MAG_IF, MAG_LF, MAG_PwrP, MAG_PwrP_R, MAG_PrinP, MAG_SampEn, MAG_ApEn, MAG_FuzEn.

These 60 input features can be divided into time domain features, frequency domain features, and nonlinear features.

Frequency domain characteristics: Frequency domain analysis in the field of signal processing can obtain more intuitive parameter characteristics than time domain analysis. In the inventive method, the frequency distribution of the signal is obtained from an energy perspective, mainly by using spectral estimation based on short signal lengths. The error caused by the coupling state is almost negligible, thereby improving the signal-to-noise ratio. Power spectral density (PSD) is broadly defined as the signal power per unit frequency band and reflects the signal power distribution in the frequency domain. In the frequency domain, the power spectrum is first calculated, and then the main tremor frequency, peak power and tremor stability index are extracted, which are all characteristics of tremor frequency. In addition, in signal processing, frequency domain analysis gives more intuitive parametric characteristics than time domain analysis. The peak power P _m (f _p ) in the signal band is defined as the area under the PSD curve in the main frequency interval, and is calculated by formula (4).

Frequency domain analysis in the field of signal processing can obtain more intuitive parameter characteristics than time domain analysis. In the inventive method, the frequency distribution of the signal is obtained from an energy perspective, mainly by using spectral estimation based on short signal lengths. The error caused by the coupling state is almost negligible, thereby improving the signal-to-noise ratio. Power spectral density (PSD) is broadly defined as the signal power per unit frequency band and reflects the signal power distribution in the frequency domain. The calculation of the power spectral density P _S (f) of the acceleration signal a _svm representing the tremor is given by Equation (3-4)

是S_dft(f)的复共轭，表示功率信号a_svm的离散傅里叶变换，可由公式(5)计算。where f _p ± f _th represents the bandwidth of the peak power at the dominant frequency. Several studies have concluded that the peak power is superior to the one-sided power spectrum of the sensor signal over a period of time (15s) around the dominant frequency ±0.5Hz.

is the complex conjugate of S _dft (f), representing the discrete Fourier transform of the power signal a _svm , which can be calculated from equation (5).

Further from equation (6), the power spectral density (PSD) P _S (f) can be calculated:

Preferably, the method of the present invention selects the Welch method, that is, after the entire acceleration signal a _svm is segmented, each small signal sequence is preprocessed and a Blackman window is added. Spectral estimation is done by a piecewise averaging periodogram method, which reduces spectral leakage.

In addition, the method of the present invention calculates the ratio of the peak power to the total power, which is used to represent the proportional relationship between the occurrence of tremor and the total recording time. The percentage of the peak power of the complete power estimate should be more significant than 85%, so that it can be determined whether the patient is in tremor or not. state. The estimated peak power of the tremor signal with respect to the PSD is shown in Fig. 3, and the shaded area represents the tremor band of the dominant frequency.

Nonlinear features:

The method of the present invention uses various entropy values to measure the complexity of tremor data. Approximate entropy (ApEn) is a technique to quantify the degree of irregularity and unpredictability of fluctuations in time series data. ApEn has the best discriminative power by comparing the frequency, RMS and ApEn's contribution to quantified tremor. Preferably, the embedding dimension is selected as m=2, the similarity tolerance is r=0.1×SD (SD is the sequence standard deviation), and ApEn is defined as

表示在相似性标准r下整个序列中所有m长度子片段的平均相似率，可计算如下：where a _svm represents a continuous 15s sequence fragment,

represents the average similarity ratio of all m-length subfragments in the entire sequence under the similarity criterion r, which can be calculated as follows:

The sample entropy (SampEn) does not include the comparison of its vectors when calculating the sequence self-similarity probability, so it is not limited by the data length. In contrast, fuzzy entropy proposes an ambiguous membership function that improves similarity measures for binary processes. This fuzzy boundary measure enhances the complexity of the signal through fuzzy entropy, making the change of entropy more continuous and smooth. Entropy features characterizing sequence complexity can substantially improve the performance of tremor quantification models, so the method of the present invention computes these nonlinear features.

Further, an automatic scoring system for tremor severity with kinematic characteristics was developed using several machine learning algorithms. Ensemble tree models, support vector machines (SVM), discriminant analysis (DA), naive Bayes and k-nearest neighbors (KNN) algorithms. The constructed SVM classifier uses three kernels (linear, polynomial and radial basis functions (RBF)). The KNN classification model uses an odd number between 1-11 as the K value. The best global solution for each model is selected using grid search and Bayesian optimization algorithms. The ensemble function of the Bayesian optimization algorithm is expected to improve every second, and each model iterates for 30 epochs to obtain the optimal solution. For models with few hyperparameters, grid search can be preferred to obtain the optimal solution. The specific classifier parameter design is shown in the following table.

Table 2 Hyperparameter settings for machine learning classifiers

The present invention defines verification conditions to ensure the generalization of the classification model.

其分类误差e_test计算如下：Training uses five-fold cross-validation to reduce the bias of the classification results. In a limited training set, five-fold cross-validation is the most appropriate validation method to train all classes without overfitting bias. The method of the present invention defines the absolute error for CRST classification to evaluate the performance of the automatic scoring system. For machine learning models

Its classification error e _test is calculated as follows:

是分类器确定的震颤等级。除最小分类误差外，本研究还计算了混淆矩阵的各个分类指数和多个分类的AUC值。where I( ) represents the indicator function, _yi is the consensus score of the three neurologists on the CRST scale of the i-th ET patient,

is the tremor level determined by the classifier. In addition to the minimum classification error, this study also calculated the individual classification indices of the confusion matrix and the AUC values for multiple classifications.

Further, the method of the present invention defines a classification model performance evaluation method to evaluate the classification performance of tremor severity in full name. Preferably, the method of the present invention uses four main indicators to evaluate the performance of arrhythmia detection classification results, including accuracy (ACC), sensitivity (SED), specificity (SPEC), precision (PRE) and F1 score, defined As follows (10-14).

TP (true positive) means that the classification is correct, and the samples that originally belonged to the positive class are divided into positive classes; TN (true negative) means that the classification is correct, and the samples that originally belonged to the negative class are divided into negative classes; FP (false positive) means that the classification is wrong , divide the errors that originally belonged to the negative class into the positive class; FN (false negative) represents the classification error, and divide the errors that originally belonged to the positive class into the negative class. Since the F1 score weights FP and FN equally, it provides a less biased metric than accuracy. In contrast, receiver operating characteristic (ROC) curves take into account classification thresholds that trade off sensitivity and specificity. The area under the curve (AUC) is often used as an evaluation metric when databases are unevenly distributed.

Further, the method of the present invention shows the classification performance of each optimized machine learning method in three postural tremor tasks through Table 4. It can be seen that the accuracy of all algorithms can be maintained above 85.72%, indicating that the classification results are in good agreement with the actual scores. However, both sensitivity and precision scores are low, so the F1 score, as a harmonic mean of sensitivity and precision, is more indicative of the ability to control both risk and decision cost. Combining the F1 scores of the three tasks, the experimental results show that AdaBoost has the best performance with an average of 93.47%. In addition, AUC, as an indicator for weighing sensitivity and specificity, can better reflect the generalization ability of the model in imbalanced datasets. According to the ROC curves in Figure 4(a-c), it can be found that regardless of the basic probability of each category, the AUC of AdaBoost can reach more than 99.1%. Compared with other machine learning models, the performance fluctuates less and the test results are relatively stable. . Figure 5(a–c) plots the optimal AdaBoost algorithm and confusion matrix in three pose tremor task classification tasks. It can be seen that the wing beat position is the most challenging task for the AdaBoost model. However, with the exception of mild tremor CRST 1-2, which had the lowest sensitivity of 83.70% and 89.83%, all other categories had indices above 90.04%. Only the arm span detection accuracy of tremor level 1 is poor, with sensitivities of 85.28%, 91.84%, 94.30% and 94.69%.

Table 3. Classification performance of each optimized machine learning method on three postural tremor tasks.

The experimental results of the algorithm of the present invention show that the integrated learning method based on the complete feature set can effectively classify the five severities of three postural tremors. In addition, the algorithm of the present invention is optimized for five machine learning models using different feature selection methods to evaluate meaningful clinical features and speed up the reasoning speed of the algorithm. Table 4 summarizes how five machine learning models optimize classification results based on PCA ranker, variance threshold-based filtering methods, mutual information-based filtering methods, Wrapper subset evaluation, and Embedded feature subsets. The global metrics AUC and accuracy show that the ensemble learning algorithm performs best on all feature subsets across all three postural tremor tasks. PCA reduces the number of features by 93.33% by selecting 95% of the explained variance and compressing the features to less than 4. The AUC of this method for wing beat, arm extension and finger position tasks were 0.870, 0.926 and 0.754, respectively. Although there is a slight loss in performance compared to other feature selection methods, it can greatly speed up model operations.

The algorithm of the present invention shows the ROC curves of the optimal ensemble learning model based on different feature subsets in the three tasks (Fig. 5(d-f)). For different tasks, the final selected features and algorithms are different. Finally, this study visually shows the confusion matrix of the optimal ensemble learning method trained based on the optimal feature selection method through Figure 5(d-f). The experimental results found that in the wing beat task, the optimal classification results of RUSBoost based on the embedded method selected a total of 16 feature parameters (73.33% of the features were eliminated), and the specificity of five severity was 97.46%- Between 99.54%. In contrast, using the AdaBoost model based on the Wrapper method in the arm extension task with only 12 feature parameters selected, the macro F1 score can reach 94.49%. Even for the severity 4 category, the accuracy, sensitivity, specificity, precision and F1 scores were 99.45%, 96.32%, 99.72%, 96.57% and 96.44%, respectively. The bagging model based on mutual information method is used in the pointing task. The model easily mistaken CRST 1 for no tremor symptoms, but other types of F1 scores were between 93.21%-94.47%.

Table 4. Predictive performance of machine learning classifiers based on different feature selection methods.

The invention introduces an automatic quantification method for postural tremor developed based on wearable equipment. Based on a rigorous experimental paradigm, the study collected postural signals for three tasks in patients diagnosed with ET. It builds a database (including high-definition video, electronic case information, and high-dimensional posture sensing signals) that is jointly and independently scored by three neurosurgeons. In the data analysis phase, the research explores SVM, KNN, NB, decision tree and ensemble learning methods to select the best hyperparameters through Bayesian optimization and grid search. Each of these machine learning algorithms is validated using a five-fold crossover method to enhance its generalization ability. Furthermore, we explore the representational ability of parameter subsets constructed using different feature selection methods for the postural tremor task and demonstrate the effectiveness of the extracted features.

Combined with the experimental results, the best model of the present invention is an integrated learning model, using the RUSBoost model based on the Embedded method in the wing stroke position task, using the AdaBoost model based on the Wrapper method in the arm extension position task, and in the finger position task. Use the Bagging algorithm based on the Mutual information method. The three postural tremor tasks have accuracies in the range of 97.25%-97.98%, AUCs of 0.980-0.997, and a small probability of classification error, which is the best we know of for automatic scoring of ET symptoms.

The model still has good AUC in predicting several classes, and has the best performance for automatic recognition of ET symptoms so far. These results suggest that the proposed method is suitable for applying standardized laboratory tests to help clinicians automatically score complex or early ET cases to aid decision-making and improve disease management efficiency.

Although the present invention has been described in detail with reference to the foregoing embodiments, for those skilled in the art, it is still possible to modify the technical solutions described in the foregoing embodiments, or to perform equivalent replacements for some of the technical features. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be included within the protection scope of the present invention.

Claims (10)

1. A postural tremor detection model, the model being based on an ensemble learning model, characterized in that:

the input features of the model are multiple; at least one input feature of the plurality of input features is obtained from triaxial acceleration data of the hand of the subject, at least one input feature is obtained from triaxial gyroscope data of the hand of the subject, at least one input feature is obtained from triaxial magnetometer data of the hand of the subject;

the model classifies tremor levels of the subject while in a predetermined posture according to the plurality of input features;

the preset postures comprise a wing-beating position, an arm-stretching position and an opposite-finger position.

2. The postural tremor detection model of claim 1, further characterized by:

when the wing positions are detected, the ensemble learning model adopts an RUSBoost model based on an embedding mode;

when the arm extension position is detected, the ensemble learning model adopts an AdaBoost model based on a Wrapper method;

when the finger positions are detected, the ensemble learning model adopts a Bagging model based on a mutual information method.

3. The postural tremor detection model of claim 2, wherein:

the plurality of input features are selected from:

ACC_Max、ACC_Min、ACC_Mean、ACC_PeakP、ACC_Arv、ACC_Var、ACC_Std、ACC_Kurt、ACC_Skew、ACC_RMS、ACC_SF、ACC_CF、ACC_IF、ACC_LF、ACC_PwrP、ACC_PwrP_R、ACC_PrinP、ACC_SampEn、ACC_ApEn、ACC_FuzEn；

GYRO_Max、GYRO_Min、GYRO_Mean、GYRO_PeakP、GYRO_Arv、GYRO_Var、GYRO_Std、GYRO_Kurt、GYRO_Skew、GYRO_RMS、GYRO_SF、GYRO_CF、GYRO_IF、GYRO_LF、GYRO_PwrP、GYRO_PwrP_R、GYRO_PrinP、GYRO_SampEn、GYRO_ApEn、GYRO_FuzEn；

MAG_Max、MAG_Min、MAG_Mean、MAG_PeakP、MAG_Arv、MAG_Var、MAG_Std、MAG_Kurt、MAG_Skew、MAG_RMS、MAG_SF、MAG_CF、MAG_IF、MAG_LF、MAG_PwrP、MAG_PwrP_R、MAG_PrinP、MAG_SampEn、MAG_ApEn、MAG_FuzEn；

wherein ACC _ Max, ACC _ Min, ACC _ Mean, ACC _ PeakP, ACC _ Arv, ACC _ Var, ACC _ Std, ACC _ Kurt, ACC _ Skaw, ACC _ RMS, ACC _ SF, ACC _ CF, ACC _ IF, ACC _ LF, GYRO _ Max, GYRO _ Min, GYRO _ Mean, GYRO _ PeakP, GYRO _ Arv, GYRO _ Var, GYRO _ Std, GYRO _ Kurt, GYRO _ Skaw, GYRO _ RMS, GYRO _ SF, GYRO _ CF, GYRO _ IF, GYRO _ LF, MAG _ Max, MAG _ Min, MAG _ Mean, MAG _ PeakP, MAG _ Arv, MAG _ Std, MAG _ rtt, KuRT, SkyO _ RMS, MAG _ SF, MAG _ LF, MAG _ Val, MAG _ LF, MAG _ Val, and MAG are the characteristics;

ACC _ Max is the maximum peak value of signals of the triaxial acceleration data; ACC _ Min is the signal minimum peak value of the triaxial acceleration data; ACC _ Mean is the signal average value of the triaxial acceleration data; ACC _ PeakP is a signal peak-to-peak amplitude of triaxial acceleration data; ACC _ Arv is the signal average rectification value of the triaxial acceleration data; ACC _ Var is the signal variance of the triaxial acceleration data; ACC _ Std is a signal standard deviation of triaxial acceleration data; ACC _ Kurt is the signal kurtosis of the triaxial acceleration data; ACC _ Skaw is the signal kurtosis of the triaxial acceleration data; ACC _ RMS is the signal root mean square of the triaxial acceleration data; ACC _ SF is a signal waveform factor of triaxial acceleration data; ACC _ CF is a signal peak factor of triaxial acceleration data; ACC _ IF is a signal pulse factor of triaxial acceleration data; ACC _ LF is a signal margin factor of triaxial acceleration data;

GYRO _ Max is the maximum peak value of the signal of the data of the three-axis gyroscope; GYRO _ Min is the minimum peak value of the data of the three-axis gyroscope; GYRO _ Mean is the signal average value of the data of the three-axis gyroscope; GYRO _ PeakP is the signal peak-to-peak amplitude of the data of the triaxial gyroscope; GYRO _ Arv is the signal average rectification value of the tri-axial gyroscope data; GYRO _ Var is the signal variance of the data of the three-axis gyroscope; GYRO _ Std is the signal standard deviation of the data of the three-axis gyroscope; GYRO _ Kurt is the signal kurtosis of the data of the three-axis gyroscope; GYRO _ Skaw is the signal kurtosis of the data of the three-axis gyroscope; GYRO _ RMS is the root mean square of the signal of the data of the three-axis gyroscope; GYRO _ SF is a signal form factor of the data of the three-axis gyroscope; GYRO _ CF is a signal peak factor of the data of the triaxial gyroscope; GYRO _ IF is a signal pulse factor of the data of the three-axis gyroscope; GYRO _ LF is a signal margin factor of the data of the three-axis gyroscope;

MAG _ Max is the maximum peak value of the signal of the three-axis magnetometer data; MAG _ Min is the minimum peak value of the three-axis magnetometer data; MAG _ Mean is the signal average value of the three-axis magnetometer data; MAG _ PeakP is the signal peak-to-peak amplitude of the data of the three-axis magnetometer; MAG _ Arv is the signal average rectification value of the three axis magnetometer data; MAG _ Var is the signal variance of the three-axis magnetometer data; MAG _ Std is the signal standard deviation of the three-axis magnetometer data; MAG _ Kurt is the signal kurtosis of the three-axis magnetometer data; MAG _ Skaew is the signal kurtosis of the three-axis magnetometer data; MAG _ RMS is the root mean square of the signal of the triaxial magnetometer data; MAG _ SF is a signal form factor of the three-axis magnetometer data; MAG _ CF is a signal peak factor of the triaxial magnetometer data; MAG _ IF is a signal pulse factor of the three-axis magnetometer data; MAG _ LF is a signal margin factor of the data of the three-axis magnetometer;

ACC _ PwrP, ACC _ PwrP _ R, ACC _ PrinP, GYRO _ PwrP _ R, GYRO _ PrinP, MAG _ PwrP _ R, MAG _ PrinP are frequency domain features;

ACC _ PwrP is the peak power of the triaxial acceleration data; ACC _ PwrP _ R is the peak power proportion of the triaxial acceleration data; ACC _ PrinnP is a main power peak value of triaxial acceleration data;

GYRO _ PwrP is the peak power of the data of the three-axis gyroscope; GYRO _ PwrP _ R is the peak power proportion of the data of the three-axis gyroscope; GYRO _ PrinP is the main power peak of the data of the triaxial gyroscope;

MAG _ PwrP is the peak power of the three-axis magnetometer data; MAG _ PwrP _ R is the peak power proportion of the three-axis magnetometer data; MAG _ PrinnP is the main peak value of the power of the data of the three-axis magnetometer;

ACC _ SampEn, ACC _ ApEn, ACC _ FuzEn, GYRO _ SampEn, GYRO _ ApEn, GYRO _ FuzEn, MAG _ SampEn, MAG _ ApEn, MAG _ FuzEn are nonlinear characteristics;

ACC _ SampEn is sample entropy of triaxial acceleration data; ACC _ ApEn is approximate entropy of triaxial acceleration data; ACC _ FuzEn is fuzzy entropy of triaxial acceleration data;

GYRO _ SampEn is sample entropy of the three-axis gyroscope data, and GYRO _ ApEn is approximate entropy of the three-axis gyroscope data; GyRO _ Fuzne is the fuzzy entropy of the three-axis gyroscope data.

4. The postural tremor detection model of claim 3, wherein:

the maximum splitting number of the RUSBoost model based on the embedding mode is 311, the learning devices are 483, and the learning rate is 0.8578; the plurality of input features are: ACC _ RMS; ACC _ Min; ACC _ Arv; ACC _ LF; GYRO _ RMS; GYRO _ Max; GYRO _ Min; GYRO-Arv; GYRO _ SF; GYRO _ LF; MAG _ FuzEn; MAG _ RMS; MAG _ Max; MAG _ Mean; MAG _ SF; MAG _ CF.

5. The postural tremor detection model of claim 3, wherein:

the maximum splitting number of the AdaBoost model based on the Wrapper method is 212, the learning devices are 45, and the learning rate is 0.9986; the plurality of input features are: ACC _ RMS; ACC _ Min; ACC _ Arv; ACC _ LF; GYRO _ Max; GYRO _ Min; GYRO-Arv; GYRO _ LF; GYRO _ RMS; MAG _ FuzEn; MAG _ RMS; MAG _ Max.

6. The postural tremor detection model of claim 3, wherein:

the maximum splitting number of the Bagging model based on the mutual information method is 174, 362 learners are provided, the learning rate is 0.8578, and 4 input features are randomly extracted from the input features; the plurality of input features are: ACC _ Max; ACC _ Mean; ACC _ SF; ACC _ CF; ACC _ IF; ACC _ LF; GYRO _ Max; GYRO _ Min; GYRO _ Mean; GYRO _ PeakP; GYRO _ LF; MAG _ RMS; MAG _ Max.

7. A gesture tremor detection algorithm is characterized in that three-axis acceleration data, three-axis gyroscope data and three-axis magnetometer data of a subject during designated actions are acquired through a three-axis accelerometer, a three-axis gyroscope and a three-axis magnetometer which are arranged on the hand of the subject; extracting a plurality of input features from the processed triaxial acceleration data, triaxial gyroscope data and triaxial magnetometer data; classifying the tremor level of the subject using the plurality of input features as the input features using the postural tremor detection model of any of claims 1-6.

8. A postural tremor detection apparatus comprising a computing unit;

the computing unit runs the postural tremor detection model of any of claims 1 to 6 to classify the tremor grade of the subject.

9. The postural tremor detection apparatus of claim 8, comprising a three-axis accelerometer, a three-axis gyroscope, a three-axis magnetometer; the three-axis accelerometer, the three-axis gyroscope and the three-axis magnetometer are arranged on the hand of the subject; under a specified action of a subject, the triaxial acceleration data are obtained through the triaxial accelerometer; the three-axis gyroscope data is obtained through the three-axis gyroscope; the three axis magnetometer data is obtained by the three axis magnetometer.

10. The postural tremor detection apparatus of claim 8, wherein:

the three-axis accelerometer, the three-axis gyroscope and the three-axis magnetometer are wearable devices, and the three-axis acceleration data, the three-axis gyroscope data and the three-axis magnetometer data are sent to the computing unit in a wireless mode.

Images