CN110522456A - A kind of WD based on deep learning trembles conditions of patients self-evaluating system - Google Patents
A kind of WD based on deep learning trembles conditions of patients self-evaluating system Download PDFInfo
- Publication number
- CN110522456A CN110522456A CN201910916656.0A CN201910916656A CN110522456A CN 110522456 A CN110522456 A CN 110522456A CN 201910916656 A CN201910916656 A CN 201910916656A CN 110522456 A CN110522456 A CN 110522456A
- Authority
- CN
- China
- Prior art keywords
- trembles
- frequency
- deep learning
- module
- patient
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Pending
Links
Classifications
-
- A—HUMAN NECESSITIES
- A61—MEDICAL OR VETERINARY SCIENCE; HYGIENE
- A61B—DIAGNOSIS; SURGERY; IDENTIFICATION
- A61B5/00—Measuring for diagnostic purposes; Identification of persons
- A61B5/103—Detecting, measuring or recording devices for testing the shape, pattern, colour, size or movement of the body or parts thereof, for diagnostic purposes
- A61B5/11—Measuring movement of the entire body or parts thereof, e.g. head or hand tremor, mobility of a limb
- A61B5/1101—Detecting tremor
-
- A—HUMAN NECESSITIES
- A61—MEDICAL OR VETERINARY SCIENCE; HYGIENE
- A61B—DIAGNOSIS; SURGERY; IDENTIFICATION
- A61B5/00—Measuring for diagnostic purposes; Identification of persons
- A61B5/40—Detecting, measuring or recording for evaluating the nervous system
- A61B5/4076—Diagnosing or monitoring particular conditions of the nervous system
- A61B5/4082—Diagnosing or monitoring movement diseases, e.g. Parkinson, Huntington or Tourette
-
- A—HUMAN NECESSITIES
- A61—MEDICAL OR VETERINARY SCIENCE; HYGIENE
- A61B—DIAGNOSIS; SURGERY; IDENTIFICATION
- A61B5/00—Measuring for diagnostic purposes; Identification of persons
- A61B5/72—Signal processing specially adapted for physiological signals or for diagnostic purposes
- A61B5/7225—Details of analog processing, e.g. isolation amplifier, gain or sensitivity adjustment, filtering, baseline or drift compensation
-
- A—HUMAN NECESSITIES
- A61—MEDICAL OR VETERINARY SCIENCE; HYGIENE
- A61B—DIAGNOSIS; SURGERY; IDENTIFICATION
- A61B5/00—Measuring for diagnostic purposes; Identification of persons
- A61B5/72—Signal processing specially adapted for physiological signals or for diagnostic purposes
- A61B5/7235—Details of waveform analysis
- A61B5/7253—Details of waveform analysis characterised by using transforms
- A61B5/7257—Details of waveform analysis characterised by using transforms using Fourier transforms
-
- A—HUMAN NECESSITIES
- A61—MEDICAL OR VETERINARY SCIENCE; HYGIENE
- A61B—DIAGNOSIS; SURGERY; IDENTIFICATION
- A61B5/00—Measuring for diagnostic purposes; Identification of persons
- A61B5/72—Signal processing specially adapted for physiological signals or for diagnostic purposes
- A61B5/7235—Details of waveform analysis
- A61B5/7264—Classification of physiological signals or data, e.g. using neural networks, statistical classifiers, expert systems or fuzzy systems
- A61B5/7267—Classification of physiological signals or data, e.g. using neural networks, statistical classifiers, expert systems or fuzzy systems involving training the classification device
Landscapes
- Health & Medical Sciences (AREA)
- Life Sciences & Earth Sciences (AREA)
- Engineering & Computer Science (AREA)
- Physics & Mathematics (AREA)
- Biomedical Technology (AREA)
- Medical Informatics (AREA)
- Physiology (AREA)
- Artificial Intelligence (AREA)
- Animal Behavior & Ethology (AREA)
- Veterinary Medicine (AREA)
- Biophysics (AREA)
- Pathology (AREA)
- General Health & Medical Sciences (AREA)
- Heart & Thoracic Surgery (AREA)
- Public Health (AREA)
- Molecular Biology (AREA)
- Surgery (AREA)
- Signal Processing (AREA)
- Psychiatry (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Neurology (AREA)
- Mathematical Physics (AREA)
- Neurosurgery (AREA)
- Power Engineering (AREA)
- Dentistry (AREA)
- Oral & Maxillofacial Surgery (AREA)
- Evolutionary Computation (AREA)
- Fuzzy Systems (AREA)
- Developmental Disabilities (AREA)
- Measurement Of The Respiration, Hearing Ability, Form, And Blood Characteristics Of Living Organisms (AREA)
Abstract
It trembles conditions of patients self-evaluating system, including signal acquisition module of trembling the present invention relates to a kind of WD based on deep learning, is used to acquire WD and trembles the signal that trembles of patient；Signal processing module is used to isolate and tremble and the non-window that trembles；Characteristic extracting module is used to extract frequency and amplitude characteristic in the signal that trembles；Feature selection module is used to select a character subset from relatively more available feature；Categorization module and grading module neural network based, the correlation to be scored by analyzing characteristic value with doctor, so that prediction scoring and the scoring of doctor have higher correlation.The present invention is collected by data of the finger inertia node to frequency and amplitude that patient trembles, to the deep learning of data, and then obtain patient and tremble the analysis of degree, the grasp convenient for doctor to curative effect is of great significance to instructing doctor to improve treatment method etc..
Description
Technical field
The present invention relates to a kind of assessment systems by deep learning, shake more particularly, to a kind of WD based on deep learning
It quivers conditions of patients self-evaluating system.
Background technique
Wilson disease (Wilson ' s disease, WD) is also known as hepatolenticular degeneration (hepatolenticular
Degeneration, HLD), it is a kind of copper dysbolism disease of autosomal recessive inheritance, caused by copper dysbolism
With the characteristics of brain degenerative disease based on cirrhosis, basal ganglia lesion.There are 330 (68.2%) in 484 HLD patients of research
It trembles, in clinical examination, the degree that patient trembles is judged often only by doctor's observation.
The performance of HLD patients' neural's system is generally present in 12~30 years old patient, average age about 19 years old, often slowly develops,
There can be interim alleviation or aggravate, also have progressed rapid person, especially young patient.Nervous system clinical manifestation outstanding is cone
It is in vitro symptom, shows as limbs dancing sample and the movement of athetoid sample, myodystony, strange expression, inactive, Intentionality
Or postural tremor, myotonia, bradykinesia, dysarthrosis, dysphagia, flexion posture and festinating gait etc..In the manhood
Common inactive and/or position property are trembled in the patient of onset, have 330 (68.2%) to go out in 484 HLD patients of research
Now tremble, what HLD occurred tremble belongs to position property and trembles mostly, or based on being trembled with position property, exist simultaneously Intentionality and/or
Static tremor also has small number of patients using intentional tremor as main physical signs.Trembling, it is unilateral or double to be limited only in morbidity early stage
Side finger and/or wrist, build up forearm, upper arm, trunk, double lower limb and incidence, part HLD also visible lower jaw, facial muscle,
Lingualis and throat tremble.The amplitude trembled is different, and it is slight that both hands only occur when two-arm is flattened forward in early stage or light-duty patient
It is tiny to tremble.If failing to obtain correct diagnosis in time and driving copper treatment, tremble while position extends, amplitude also gradually adds
Greatly.Typical critically ill patient, no matter in static or active procedure, presentation wrist, both arms, which are hit, claps sample movement, so that oneself cannot
Self-care diet, the daily lifes such as wear the clothes.Light-duty patient tremble often can by the will of short duration control several seconds to several minutes, but it is out of control after
Tremor amplitude transient can aggravate；Severe tremor is not controlled generally by will.For the classification trembled, most HLD suffer from
Person shows as position property and trembles and intentional tremor.
2006, depth structure chemistry was practised, or was usually more known as deep learning or Layered Learning, had become machine
The frontier of device Learning Studies.2006, University of Toronto professor, machine learning field authority Geoffrey
Hinton publishes thesis in " science " proposes deep learning main points of view:
(1) artificial neural network of more hidden layers has excellent feature learning ability, and the feature learnt has data
More essential portrays, to be conducive to visualize or classify.
(2) difficulty of the deep neural network in training can pass through " successively initialization " (layer-wise pre-
Training) Lai Youxiao overcomes, what successively initialization can be realized by unsupervised learning.In past this several years, depth
The development technique of habit produces wide influence to traditional Signal and Information Processing research, widely from the point of view of, also include machine
The key areas such as device study and artificial intelligence, see (summary articles such as Bengio, 2013；Hinton etc., 2012；Yu and Deng,
2011；Deng, 2011；Arel etc., 2010, the New York Times also reported in media report this progress (Markoff,
2012).A series of nearest seminars, monograph, operating room etc. all support the depth of investigation to learn and in Signal and Information Processing
The various applications in field.
The learning methods such as Most current classification, recurrence are shallow structure algorithm, are limited in that finite sample and calculating
Limited to the expression ability of complicated function under cell cases, for complicated classification problem, its generalization ability is centainly restricted.It is deep
Degree study can realize that complicated function approaches, characterize input data distribution table by learning a kind of deep layer nonlinear network structure
Show, and presents the powerful ability from a few sample focusing study data set substantive characteristics.(benefit of multilayer be can with compared with
Few parameter list is given instructions in reply miscellaneous function).
Deep learning can obtain the feature for preferably indicating data, simultaneously because the level of model, parameter are very
More, capacity is enough, and therefore, model has the ability to indicate large-scale data, so unobvious for image, this feature of voice (need
Want hand-designed and much without intuitive physical meaning) the problem of, better effect can be obtained on large scale training data.
In addition, the angle of slave pattern identification feature and classifier, feature and classifier are integrated to a frame by deep learning frame
In frame, learning characteristic is removed with data, reducing the huge workload of hand-designed feature in use, (this is current industry work
Cheng Shi works hard most aspects), therefore, not only effect can be more preferable, but also uses also and have many conveniences, is
A set of frame that machine learning field extremely merits attention.
Summary of the invention
The present invention devises a kind of WD based on deep learning and trembles conditions of patients self-evaluating system, and the technology solved is asked
Topic can only be measured, can not be obtained by doctor's range estimation and subjective judgement to the detection of patient trembled during Present clinical checks
More objective out and accurate detection data.
In order to solve above-mentioned technical problem, present invention employs following scheme:
A kind of WD based on deep learning trembles conditions of patients self-evaluating system, and feature exists: including signal acquisition of trembling
Module is used to acquire WD and trembles the signal that trembles of patient；Signal processing module is used to isolate and tremble and the non-window that trembles
Mouthful；Characteristic extracting module is used to extract frequency and amplitude characteristic in the signal that trembles；Feature selection module is used for from phase
To one character subset of selection in more available feature；Categorization module and grading module neural network based, pass through
The correlation that analysis characteristic value scores with doctor, so that prediction scoring and the scoring of doctor have higher correlation.
Preferably, the experimental facilities that signal acquisition uses that trembles in the signal acquisition module of trembling is finger inertia section
Point, using nine axle sensor of MPU6050, including 3 axis accelerometers, 3 axis gyroscopes and 3 axis magnetometers, acquisition is trembled respectively
Acceleration information, gyro data and magnetic field data.
It preferably, further include filtering in the signal acquisition module of trembling, MPU6050 further includes built-in amplifier, low pass
Filter and 16 ADC directly can export digital quantity by I2C bus, be not necessarily to external conversion and filter circuit.
Preferably, No. 1 fingerstall of the finger inertia node and No. 2 fingerstall press label cover on thumb and index finger, experiment
First is that test static tremor, allows the patient for wearing equipment to be sitting on chair, both hands are placed on thigh and remain stationary, and wait patients
Stable posture after start save data, test 30 seconds more than after stopping；Test two postural tremors and experiment three flapping wings shake
Quivering is that test arm trembles, and patient's both arms under standing state is allowed to stretch forward first, is kept with shoulder with width, followed by it is flat to bend elbow
It is put in front, pronation, the two experiments are respectively to acquire 30 seconds above data after waiting patients' stable posture；If it is unknown to tremble
It is aobvious, survey can be added to bend elbow and lay flat the movement that front keeps palm upward.
Preferably, it is subsequent also it is achievable experiment four, mediate and clench fist, corresponding amount list item: finger mediate (B15)；Firstly, allowing
Patient stretches out thumb and index finger, alternately carries out opening and closing movement, and frequency is as fast as possible, and amplitude is as big as possible, save 20 times or more
Data；Then, the movement for allowing patient to be alternately performed the five fingers stretching, extension and hold with a firm grip, equally also saves 20 above data.
Preferably, subsequent there is also experiment fives, hand alternating movement, and corresponding amount list item: both hands answer alternate motion fastly
(B16)；Sensor is installed: body node, at bilateral wrist；Design action: palm, the back of the hand of patient when standing with a hand are allowed
The palm for alternately patting another hand, is repeated 20 times above and saves data.
Preferably, windowing process is carried out to signal using Short Time Fourier Transform (STFT) in the signal processing module,
It isolates and trembles and the non-window that trembles.
Preferably, frequency analysis is carried out by extracting fundamental frequency in the characteristic extracting module, it is right again by frequency analysis
Type of trembling is classified；
Power spectral density function, the function representation are calculated using tremor amplitude as classified adaptive factor in the characteristic extracting module
Signal power under different frequency；This it appears that the dominant frequency trembled, average to shake from the visible peak of power spectral density (PSD)
Flutter amplitude degree is determined by peak value region below；In order to obtain the range of PSD, to each window of patient's acceleration signal into
A series of processing of row, obtain exact numerical, on the basis of formula (1), calculate the PSD weighted average absolute value of each frequency range:
(1) Fst represents the initial frequency of each frequency range in formula, and Fsp is off frequency, and i represents the Fst based on sample frequency
Frequency step between Fsp, P and f are power corresponding with i value and frequency.
Preferably, the feature quantity l that feature selecting is chosen in the feature selection module is less than original characteristic point M, subtracts
Few characteristic point quantity, dimensionality reduction avoid the over-fitting to specific training dataset, design the classification with good Generalization Capability
Device；By Brute-force search algorithm, various possible combinations are formed, calculate the classification separability of every kind of combination.
Preferably, the purpose of training multilayer perceptron is wrapped in estimation network in the categorization module neural network based
The weight and threshold value of all neurons contained select (2) formula as error function；Wherein, xiTo input,Indicate the defeated of network
Out, target is to calculate unknown weight:
Keep J minimum using backpropagation (BP) algorithm herein, is selected since different initial values, after operation repeatedly best
Solve corresponding weight；In back-propagation algorithm, estimate that weight w's in each step interative computation is previous by formula (3)
Value and current value；
W (new)=w (old)+Δ w (3)
Error term is related with the gradient of cost function, is calculated with the preceding value of w:
W refers to the weight parameter of network neural member；The performance of algorithm is heavily dependent on the value of learning rate μ, is
Guarantee convergence, the value of μ should be sufficiently small, but cannot be too small, if too small, to may result in convergence rate non-
It is often slow；In the case, introduce momentum d, can better control error rate, solve the above problems；Pass through formula (5) Lai Tigao
Learning rate:
The quantity of weight should be lacked as far as possible in network, so that the value of error function is reduced to suitable level, learn simultaneously
The quantity for practising sample should be as more as possible.The WD based on deep learning trembles conditions of patients self-evaluating system with below beneficial to effect
Fruit:
(1) present invention is collected by data of the finger inertia node to frequency and amplitude that patient trembles, to data
Deep learning (Deep Learning), and then obtain patient and tremble the analysis of degree, the grasp convenient for doctor to curative effect is right
It instructs doctor to improve treatment method etc. to be of great significance.
(2) present invention combines medical portable equipment with artificial intelligence, and realization, which is seen a doctor, to be not required to be admitted to hospital, and doctor remotely refers to
It leads Patient drug to take, treats the novel diagnosis and treatment mode of autonomy-oriented.
(3) present invention acquisition mass data, the classification and identification algorithm of the unsupervised deep learning of proposition, accuracy can be into one
Step improves, and under the mass data of various diseases, hepatolenticular degeneration patient and normal person can be preferably identified from data
Between difference, it is subsequent to extend to other the nervous system diseases, by they it is correct classify, realize the machine automatic identification cause of disease.
Detailed description of the invention
Fig. 1: the process frame diagram for signal acquisition of trembling in the present invention；
Fig. 2: tremble test action schematic diagram in the present invention；
Fig. 3: hepatolenticular degeneration is trembled the supervised learning process figure of assessment in the present invention；
Fig. 4: WD subject and the frequency domain comparison diagram for perfecting subject's left hand and right hand data in the present invention；
Fig. 5: WD patient's figure compared with the frequency domain of Healthy People in the present invention；
Fig. 6: the power spectrum chart of WD patient and healthy person in the present invention；
Fig. 7: WD subject and health volunteer's comparative bid parameter in the present invention；
Fig. 8: the effect picture classified using ANN to mean value in the present invention.
Specific embodiment
Below with reference to Fig. 1 to Fig. 8, the present invention will be further described:
1, it trembles the acquisition and filtering of signal；
Screening hepatolenticular degeneration first trembles the more apparent patient of symptom as subject, and exclusion can not independent ambulation
Or it is unable to complete paleocinetic patient, while also to exclude to lead to the patient of dyskinesia disease with other.
The experimental facilities that signal acquisition uses is finger inertia node, is added using nine axle sensor of MPU6050, including 3 axis
Speedometer, 3 axis gyroscopes and 3 axis magnetometers obtain acceleration information, gyro data and magnetic field data, MPU6050 respectively
Including built-in amplifier, low-pass filter and 16 ADC, digital quantity directly can be exported by I2C bus, without external conversion and
Filter circuit.
In order to save experimental period, avoid patient tired, it is necessary to which the sequencing of reasonable arrangement experiment is reduced to greatest extent
Unnecessary equipment handling and movement repeat.Specific experiment process is as shown in Figure 1.Data are transmitted using wifi WLAN.
Finger inertia node is to tie up sensor module on finger by velcro, and wireless transmitter module is tied up in wrist, in this way may be used
To keep globality, unnecessary vibration is avoided.
Corresponding amount list item: static tremor (B12), arm tremor (B18).
Sensor installation: No. 1 of finger node and No. 2 fingerstall press label cover on thumb and index finger, it is to note that have sensor
One side upward.It is unilateral or bilateral, it be according to the demand decision of tremble situation and the follow-up study of patient.
Design action: 3 movements are shared.First (experiment one) is test static tremor, allows the disease for wearing equipment
People is sitting on chair, and both hands are placed on thigh and remain stationary, and starts to save data after waiting the stable posture of patients, after test 30 seconds
Stop.Second is that test arm trembles with third movement, allows patient's both arms under standing state to stretch forward first, keeps
With shoulder with width, followed by bends elbow and lie against front, pronation.The two movements are also respectively to acquire 30 after waiting patients' stable posture
Second data.If it is unobvious to tremble, survey can be added to bend elbow and lay flat the movement that front keeps palm upward, the schematic diagram of everything is such as
Shown in Fig. 2:
What the present invention mainly studied is the Classification and Identification that HLD trembles, corresponding UWDRS scale, subsequent also achievable experiment four,
It mediates and clenches fist, corresponding amount list item: finger mediates (B15).
Equipment installation: finger node, thumb and index finger respectively wear a fingerstall.
Design action: firstly, patient is allowed to stretch out thumb and index finger, opening and closing movement is alternately carried out, frequency is as fast as possible, width
Spend data as big as possible, preservation is 20 times.Then, the movement for allowing patient to be alternately performed the five fingers stretching, extension and hold with a firm grip, equally also saves
20 data.
Experiment five, hand alternating movement, corresponding amount list item: both hands answer alternate motion (B16) fastly.
Sensor is installed: body node, at bilateral wrist.
Design action: it allows patient alternately to pat the palm of another hand when standing with the palm of a hand, the back of the hand, repeats 20
It is secondary and save data.
For the above experiment, a large amount of signal datas that tremble of acquisition use point of the unsupervised deep learning proposed in experiment
Class recognizer, accuracy can further improve, and under the mass data of various diseases, liver can be preferably identified from data
Difference between lenticular degeneration patient and normal person, it is subsequent to extend to other the nervous system diseases, they are correctly classified,
Realize the machine automatic identification cause of disease.
2, signal processing；
The signal that trembles includes many non-stationaries or short-time characteristic, such as drift, trend and mutation.These are generally characterized by trembling
Most important part in equal signals, therefore the present invention carries out windowing process to signal using Short Time Fourier Transform (STFT), point
It separates out and trembles and the non-window that trembles.The Hamming window of STFT has a 4 seconds length and 50% overlapping, these features of STFT into
One step attempt separation tremble with it is non-tremble window when obtain optimum.
3, feature extraction；
What the research discovery of early stage was trembled is mainly characterized by frequency and intensity, the two features are suitable for various types of shakes
It quivers.It is periodic swinging that hand, which trembles, and it is the basis of classification of trembling that fundamental frequency, which extracts, therefore the present invention is carried out by extracting fundamental frequency
Frequency analysis again classifies to type of trembling by frequency analysis.It, can be with according to the frequency detected in signal processing module
Distinguish basic type of trembling.On the basis of segmenting in front, using filter block RT (3-6Hz), PT (6-9Hz) is to the base to tremble
This type is classified.
Second essential characteristic that hand trembles is amplitude.On the basis of the studies above, using tremor amplitude as classification
The thought of the factor is the signal power in order to calculate power spectral density function, under the function representation different frequency.From power spectrum
It spends on the visible peak of (PSD) this it appears that the dominant frequency trembled, average tremor amplitude are determined by peak value region below.For
The range for obtaining PSD carries out a series of processing to each window of patient's acceleration signal, obtains exact numerical, in public affairs
On the basis of formula (1), the PSD weighted average absolute value of each frequency range is calculated:
(1) Fst represents the initial frequency of each frequency range in formula, such as the initial frequency of RT frequency range is 3Hz, then Fst=3,
Fsp is off frequency, and corresponding RT frequency range is 6Hz.I represents the frequency step between Fst and Fsp based on sample frequency.P and
F is power corresponding with i value and frequency.
It is also needed with correlation analysis using SF50 frequency and F50 frequency to preferably be compared.
One half-power frequency of F50 frequency the left side the other half on the right, power can be reasonably assigned to frequency band by it
It is interior.What SF50 was indicated is the frequency band for accounting for total power signal 68%.SF50 is centered on F50, it can be said that SF50 is represented point
The frequency being dispersed near centre frequency.The F0 of some patients, fundamental frequency, the value of F50 be not identical, therefore their difference can be made
It is characterized.
4, feature selecting；
One character subset is selected from relatively large number of available feature using the best way, these features include to understand
Certainly main information required for the classification problem of next step.Another main points of feature selecting are that the feature quantity l chosen is less than
Original characteristic point M.Characteristic point quantity is reduced, dimensionality reduction avoids the over-fitting to specific training dataset, and designing has well
The classifier of Generalization Capability, that is, when all data of test data set are all not included within training set, classifier
Operational effect is still fine.
By Brute-force search algorithm, various possible combinations are formed, calculate the classification separability of every kind of combination.
In exhaustive type classifier, for comparison result, the classification of clear classifier is needed, it is clear that due to the block and divide
Feedback link between class device block, this technique improves the precision of classifier, but also increase calculating cost.However calculate cost
It is not important on this problem, because the realization of this block is to increase precision by reducing dimension.
5, classification neural network based；
The reliability and accuracy for choosing whether rationally to determine system of sorting algorithm.Exist in neural network from output
To the feedback link of context node, it is contemplated that input the dimension of classifier, the present invention has selected artificial neural network (ANN) side
Method, this method can preferably reflecting value robustness variation, classification accuracy is higher.
The purpose of training multilayer perceptron is to estimate the weight and threshold value of all neurons for including in network.For this purpose, choosing
(2) formula is selected as error function.Wherein, xiTo input,Indicate the output of network, target is to calculate unknown weight.
Keep J minimum using backpropagation (BP) algorithm herein, is selected since different initial values, after operation repeatedly best
Solve corresponding weight.In back-propagation algorithm, estimate that weight w's in each step interative computation is previous by formula (3)
Value and current value.
W (new)=w (old)+Δ w (3)
Error term is related with the gradient of cost function, is calculated with the preceding value of w:
W refers to the weight parameter of network neural member.The performance of algorithm is heavily dependent on the value of learning rate μ, is
Guarantee convergence, the value of μ should be sufficiently small, but cannot be too small, if too small, to may result in convergence rate non-
It is often slow.In the case, introduce momentum a, can better control error rate, solve the above problems.Pass through formula (5) Lai Tigao
Learning rate.
The research of early stage has shown that the quantity of weight should be lacked as far as possible in network, so that the value of error function is reduced to conjunction
Suitable level, while the quantity of learning sample should be as more as possible.
By experiment and data processing, Fig. 4 is any one WD subject and any one name of control group in choice experiment group
Perfect right-hand man's static tremor waveform diagram of subject.Wherein, WD subject's left hand trembles scoring for 3, and right hand scoring is 1,
Perfect subject due to atremia, assert that right-hand man's scoring is 0.
It compares WD subject and perfects subject, observe that the amplitude frequency diagram of WD subject has apparent peak between 3~8Hz
Value, and amplitude is larger, and perfect subject without apparent peak value, and amplitude is smaller, more gently.It is tested to compare WD
The left hand and right hand amplitude frequency diagram of person observes that the peak value of left hand is bigger than the peak value of the right hand, and its crest frequency is also different.Explanation
Frequency domain data can quantify severity of trembling.
Subject perfected to 4 WD subjects and 4 dress the obtained data of acquisition instrument and be filtered, analysis time domain and
Frequency domain character, and calculate characteristic value.Include 8 subjects, 5 kinds of characteristic values, MEAN (mean value), RMS (root-mean-square value), pf (peak value
Frequency), pm (peak amplitude) and Ppeak (peak power), analyze between 5 kinds of characteristic values and subject's static tremor scoring
Related coefficient, reflect characteristic value and scoring between correlation.As can be seen from the figure MEAN, RMS, Pm and Ppeak
Good relationship between scoring, related coefficient are respectively r=0.886, r=0.885,0.883, and 0.816, pass through research hair
The logarithm of these four existing characteristic values and the correlation of scoring are higher, respectively r=0.97, and 0.971,0.954,0.963.And pf
It is r=0.093 with the related coefficient of scoring, the degree of correlation of the two is extremely weak, can be considered as uncorrelated, shows that crest frequency is shake
The characteristic quivered, it is not related with the severity of symptom.
Analyze the feature of data on frequency domain after carrying out DFT.Fig. 5 is pair of WD subject and health volunteer on frequency domain
Than.Observe that the amplitude frequency diagram of WD patient has significant peak value between 3-8Hz, and amplitude is very big, and Healthy People is then without apparent
Peak value, amplitude are smaller and relatively gentle.
It can preferably reflect the energy of signal with the peak power that psd (formula 1) calculates, and have well with scoring of trembling
Correlation.Total peak power is the sum of each axle acceleration signal peak power.Fig. 6 shows WD patient and normal person's power spectrum
Compare.
The data of 43 WD patients and 9 physical examination of healthy population are filtered, time domain and frequency domain character are analyzed, are calculated special
Sign, obtains data.Fig. 7 has selected being averaged for 2 subjects (1 WD subject, 1 health volunteer) from 52 subjects
5 features such as value, RMS, PF (crest frequency), PM (peak amplitude) and Ppeak (peak power).Analyze 5 characteristic values with by
Related coefficient between examination person's static tremor scoring, with the correlation reflected between characteristic value and scoring.It can from Fig. 7
Out, average value, the peak RMS, PM and P and score have good correlation, and correlation coefficient r is respectively 0.891,0.867,0.881 and
0.816.It can be seen from figure 7 that the good relationship of mean value, root mean square, the peak PM and P and scoring, correlation coefficient r are respectively
0.891,0.867,0.881,0.816.Statistics indicate that correlation coefficient r=0.128 of Pf, the degree of association of the two is not high, can see
Make without too big association, crest frequency is the feature trembled.It is unrelated with the severity of symptom.
Using different patients tremble assessment feature set as the input vector of neural network classifier, input vector it is initial
Dimension is equal to the number of extracted feature, and the quantity of output neuron is set as 5, the i.e. quantity in identification class, and each class is by defeated
The unit value signal of neuron indicates out.
The Application of Neural Network S-shaped neuron of two layers of hidden layer is adjusted hidden under search modes with heuritic approach
The number of neuron is hidden to obtain highest recall precision.Hidden neuron quantity excessively will lead to network generalization decline,
Cross that few then to will lead to learning process invalid error rate excessively high.Least hidden neuron is selected by the study of heterogeneous networks, is adopted
With heuritic approach, error function is down to satisfied level.All data trembled are divided into study and test two parts,
Data are obtained in the investigation of the hepatolenticular degeneration patient to Anhui University of Chinese Medecine's Neurology Research Institute, from 43 liver beans
Shape nuclear degeneration, which is trembled, extracts 787 experiment patterns in patient for training, and separately has 201 of 9 non-hepatolenticular degeneration patients
Training template.
It is trained using different number of hidden neuron and finds out optimal network structure, test data selects training number
According to the data except collection.The present invention is extracted using hidden neuron as few as possible and assemblage characteristic, selects test error minimum
Conduct optimal value.It is higher that the reduction of input quantity means that ability is more easily recognized in neural network structure.It is individual in feature set
The very high feature of recognition capability is not necessarily best.Therefore, the process neural network that training and test obtain again is complicated
Degree reduces.Best feature set has optimal identification, and Fig. 7 lists the efficiency parameters of categorizing system.Final nerve net
Network system has 4 input parameters, 2 layers of hidden layer with 5 S-shaped neurons.Fig. 8 shows being averaged for artificial neural network
Classification results, accuracy rate 92%.
The present invention devises a kind of wearable limbs end and trembles data acquisition device, for measuring the hand of disturbances in patients with Parkinson disease
Tremble acceleration signal.Time-domain analysis carried out to the digital signal trembled of acquisition and discrete Fourier transform is analyzed and trembled
Characteristic parameter.By analyzing the correlation that scores with doctor of characteristic value, four characteristic parameters (mean value, root mean square, the peak PM, P)
Value and scoring highly relevant (r > 0.8), prediction scoring and the scoring of doctor have higher correlation.Most probable peanut it is hidden
Neuroid and the neural network for extracting feature, select test error the smallest as optimum value.Input the reduction meaning of quantity
Taste the simplification of neural network structure and the raising of recognition capability.Feature with good independent distinguishing ability, with other spies
It is not necessarily best when sign is in identity set.The system is suitable for the diagnosis and long-range control of neurodegenerative movement obstacle
System, is especially voluntarily in the early stage of disease and patient and diagnoses.The system of design will make the family of WD patient
Monitoring is possibly realized, and neurosurgeon can change drug therapy according to long-term observation rather than simple outpatient service.
Above in conjunction with attached drawing, an exemplary description of the invention, it is clear that realization of the invention is not by aforesaid way
Limitation, as long as use the inventive concept and technical scheme of the present invention carry out various improvement, or it is not improved will be of the invention
Conception and technical scheme directly apply to other occasions, be within the scope of the invention.
Claims (10)
- The conditions of patients self-evaluating system 1. a kind of WD based on deep learning trembles, feature exist: including the signal acquisition mould that trembles Block is used to acquire WD and trembles the signal that trembles of patient；Signal processing module is used to isolate and tremble and the non-window that trembles； Characteristic extracting module is used to extract frequency and amplitude characteristic in the signal that trembles；Feature selection module is used for from relatively more Available feature in select a character subset；Categorization module and grading module neural network based, it is special by analysis The correlation that value indicative scores with doctor, so that prediction scoring and the scoring of doctor have higher correlation.
- The conditions of patients self-evaluating system 2. the WD according to claim 1 based on deep learning trembles, it is characterised in that: institute Stating the experimental facilities that signal acquisition uses that trembles in signal acquisition module of trembling is finger inertia node, using nine axis of MPU6050 Sensor, including 3 axis accelerometers, 3 axis gyroscopes and 3 axis magnetometers, obtain the acceleration information to tremble, gyroscope number respectively Accordingly and magnetic field data.
- The conditions of patients self-evaluating system 3. the WD according to claim 2 based on deep learning trembles, it is characterised in that: institute Stating in signal acquisition module of trembling further includes filtering, and MPU6050 further includes built-in amplifier, low-pass filter and 16 ADC, can Digital quantity is directly exported by I2C bus, is not necessarily to external conversion and filter circuit.
- The conditions of patients self-evaluating system 4. the WD according to claim 2 or 3 based on deep learning trembles, feature exist In: No. 1 fingerstall and No. 2 fingerstall of the finger inertia node are by label cover on thumb and index finger, and experiment is first is that test is static Property is trembled, and allows the patient for wearing equipment to be sitting on chair, both hands are placed on thigh and remain stationary, after the stable posture for waiting patients Start to save data, stop after test 30 seconds or more；Testing two postural tremors and three asterixis of experiment is test arm It trembles, patient's both arms under standing state is allowed to stretch forward first, keep with shoulder with width, followed by bend elbow and lie against front, hand Downwards, the two experiments are respectively to acquire 30 seconds above data after waiting patients' stable posture to the palm；If it is unobvious to tremble, survey can be added Elbow in the wrong lays flat the movement that front keeps palm upward.
- The conditions of patients self-evaluating system 5. the WD according to claim 4 based on deep learning trembles, it is characterised in that: after Continuous also achievable experiment four is mediated and is clenched fist, and corresponding amount list item: finger mediates (B15)；Firstly, patient is allowed to stretch out thumb and food Refer to, alternately carry out opening and closing movement, frequency is as fast as possible, and amplitude is as big as possible, preservation 20 times or more data；Then, disease is allowed The movement that people is alternately performed the five fingers stretching, extension and holds with a firm grip, equally also saves 20 above data.
- The conditions of patients self-evaluating system 6. the WD according to claim 5 based on deep learning trembles, it is characterised in that: after Continuous there is also experiment fives, hand alternating movement, and corresponding amount list item: both hands answer alternate motion (B16) fastly；Sensor installation: body Node, at bilateral wrist；Design action: patient is allowed alternately to pat the hand of another hand when standing with the palm of a hand, the back of the hand The palm, is repeated 20 times above and saves data.
- 7. the WD described in any one of -6 based on deep learning trembles conditions of patients self-evaluating system according to claim 1, It is characterized by: carrying out windowing process, separation to signal using Short Time Fourier Transform (STFT) in the signal processing module It trembles out and the non-window that trembles.
- 8. the WD described in any one of -7 based on deep learning trembles conditions of patients self-evaluating system according to claim 1, It is characterized by:Frequency analysis is carried out by extracting fundamental frequency in the characteristic extracting module, type of trembling is carried out again by frequency analysis Classification；Power spectral density function is calculated using tremor amplitude as classified adaptive factor in the characteristic extracting module, the function representation is different Signal power under frequency；This it appears that the dominant frequency trembled, averagely tremble width from the visible peak of power spectral density (PSD) Degree is determined by peak value region below；In order to obtain the range of PSD, one is carried out to each window of patient's acceleration signal Series of processes obtains exact numerical, on the basis of formula (1), calculates the PSD weighted average absolute value of each frequency range:(1) Fst represents the initial frequency of each frequency range in formula, and Fsp is off frequency, i represent Fst based on sample frequency and Frequency step between Fsp, P and f are power corresponding with i value and frequency.
- 9. the WD described in any one of -8 based on deep learning trembles conditions of patients self-evaluating system according to claim 1, It is characterized by: the feature quantity l that feature selecting is chosen in the feature selection module is less than original characteristic point M, reduce special Sign point quantity, dimensionality reduction avoid the over-fitting to specific training dataset, design the classifier with good Generalization Capability；It is logical Brute-force search algorithm is crossed, various possible combinations are formed, calculates the classification separability of every kind of combination.
- 10. the WD described in any one of -9 based on deep learning trembles conditions of patients self-evaluating system according to claim 1, It is characterized by: including in the purpose estimation network of training multilayer perceptron in the categorization module neural network based The weight and threshold value of all neurons select (2) formula as error function；Wherein, xiTo input,Indicate the output of network, Target is to calculate unknown weight:Keep J minimum using backpropagation (BP) algorithm herein, since different initial values, optimum solution pair is selected after operation repeatedly The weight answered；In back-propagation algorithm, estimated by formula (3) in each step interative computation the previous value of weight w and Current value；W (new)=w (old)+Δ w (3)Error term is related with the gradient of cost function, is calculated with the preceding value of w:W refers to the weight parameter of network neural member；The performance of algorithm is heavily dependent on the value of learning rate μ, in order to protect Convergence is demonstrate,proved, the value of μ should be sufficiently small, but cannot be too small, and to may result in convergence rate very slow if too small； In the case, introduce momentum α, can better control error rate, solve the above problems；Study is improved by formula (5) Rate:The quantity of weight should be lacked as far as possible in network, so that the value of error function is reduced to suitable level, while learn sample This quantity should be as more as possible.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201910916656.0A CN110522456A (en) | 2019-09-26 | 2019-09-26 | A kind of WD based on deep learning trembles conditions of patients self-evaluating system |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201910916656.0A CN110522456A (en) | 2019-09-26 | 2019-09-26 | A kind of WD based on deep learning trembles conditions of patients self-evaluating system |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN110522456A true CN110522456A (en) | 2019-12-03 |
Family
ID=68670255
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201910916656.0A Pending CN110522456A (en) | 2019-09-26 | 2019-09-26 | A kind of WD based on deep learning trembles conditions of patients self-evaluating system |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN110522456A (en) |
Cited By (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN111990967A (en) * | 2020-07-02 | 2020-11-27 | 北京理工大学 | Gait-based Parkinson disease recognition system |
| CN112674762A (en) * | 2020-12-28 | 2021-04-20 | 江苏省省级机关医院 | Parkinson tremble evaluation device based on wearable inertial sensor |
| CN113616194A (en) * | 2021-08-05 | 2021-11-09 | 苏州小蓝医疗科技有限公司 | Device and method for monitoring hand tremor frequency and intensity |
| CN113925495A (en) * | 2021-10-20 | 2022-01-14 | 福建工程学院 | Arteriovenous fistula abnormal tremor signal identification system and method combining statistical learning and time-frequency analysis |
Citations (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN104398263A (en) * | 2014-12-25 | 2015-03-11 | 中国科学院合肥物质科学研究院 | Method for quantitatively evaluating symptoms of tremor of patient with Parkinson's disease according to approximate entropy and cross approximate entropy |
| CN104434129A (en) * | 2014-12-25 | 2015-03-25 | 中国科学院合肥物质科学研究院 | Quantization evaluating device and method for dyskinesia symptoms of Parkinson and related extrapyramidal diseases |
| CN104660717A (en) * | 2015-03-16 | 2015-05-27 | 北京品驰医疗设备有限公司 | Working method of remote monitoring system of implantable medical device |
| CN109276255A (en) * | 2018-11-27 | 2019-01-29 | 平安科技（深圳）有限公司 | A kind of limb tremor detection method and device |
-
2019
- 2019-09-26 CN CN201910916656.0A patent/CN110522456A/en active Pending
Patent Citations (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN104398263A (en) * | 2014-12-25 | 2015-03-11 | 中国科学院合肥物质科学研究院 | Method for quantitatively evaluating symptoms of tremor of patient with Parkinson's disease according to approximate entropy and cross approximate entropy |
| CN104434129A (en) * | 2014-12-25 | 2015-03-25 | 中国科学院合肥物质科学研究院 | Quantization evaluating device and method for dyskinesia symptoms of Parkinson and related extrapyramidal diseases |
| CN104660717A (en) * | 2015-03-16 | 2015-05-27 | 北京品驰医疗设备有限公司 | Working method of remote monitoring system of implantable medical device |
| CN109276255A (en) * | 2018-11-27 | 2019-01-29 | 平安科技（深圳）有限公司 | A kind of limb tremor detection method and device |
Non-Patent Citations (1)
| Title |
|---|
| DU WEI: "analysis and classification of tremor characteristics of hepatolenticular degeneration", 《INTERNATIONAL CONFERENCE ON APPLICATIONS AND TECHNIQUES IN CYBER INTELLIGENCE ATCI 2019》 * |
Cited By (6)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN111990967A (en) * | 2020-07-02 | 2020-11-27 | 北京理工大学 | Gait-based Parkinson disease recognition system |
| CN112674762A (en) * | 2020-12-28 | 2021-04-20 | 江苏省省级机关医院 | Parkinson tremble evaluation device based on wearable inertial sensor |
| CN113616194A (en) * | 2021-08-05 | 2021-11-09 | 苏州小蓝医疗科技有限公司 | Device and method for monitoring hand tremor frequency and intensity |
| CN113616194B (en) * | 2021-08-05 | 2023-10-03 | 苏州小蓝医疗科技有限公司 | Device and method for monitoring hand tremor frequency and intensity |
| CN113925495A (en) * | 2021-10-20 | 2022-01-14 | 福建工程学院 | Arteriovenous fistula abnormal tremor signal identification system and method combining statistical learning and time-frequency analysis |
| CN113925495B (en) * | 2021-10-20 | 2023-04-21 | 福建工程学院 | Arterial and venous fistula abnormal tremor signal identification system and method combining statistical learning and time-frequency analysis |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN110522456A (en) | A kind of WD based on deep learning trembles conditions of patients self-evaluating system | |
| Pogorelc et al. | Automatic recognition of gait-related health problems in the elderly using machine learning | |
| CN109620223A (en) | A kind of rehabilitation of stroke patients system brain-computer interface key technology method | |
| Burns et al. | Upper limb movement classification via electromyographic signals and an enhanced probabilistic network | |
| Yalcin et al. | Epilepsy diagnosis using artificial neural network learned by PSO | |
| Liparulo et al. | A novel fuzzy approach for automatic Brunnstrom stage classification using surface electromyography | |
| Chang et al. | Emotion recognition with consideration of facial expression and physiological signals | |
| Krassnig et al. | User-friendly system for recognition of activities with an accelerometer | |
| Jia et al. | Human daily activity recognition by fusing accelerometer and multi-lead ECG data | |
| Ortega-Anderez et al. | A multi-level refinement approach towards the classification of quotidian activities using accelerometer data | |
| Benazzouz et al. | EMG Feature selection for diagnosis of neuromuscular disorders | |
| Naik et al. | Classification of finger extension and flexion of EMG and Cyberglove data with modified ICA weight matrix | |
| Tran et al. | Robust classification of functional and nonfunctional arm movement after stroke using a single wrist-worn sensor device | |
| CN110522455A (en) | A kind of WD based on deep learning trembles grade appraisal procedure | |
| Zhang et al. | Diagnosing Parkinson's disease with speech signal based on convolutional neural network | |
| Shair et al. | Finger movement discrimination of EMG signals towards improved prosthetic control using TFD | |
| Oikonomou et al. | Sparse kernel machines for motor imagery eeg classification | |
| CN112674760A (en) | Wearable sensor-based Parkinson upper limb movement detection method | |
| CN110321856B (en) | Time-frequency multi-scale divergence CSP brain-computer interface method and device | |
| Miller et al. | Comparison of machine learning approaches for classifying upper extremity tasks in individuals post-stroke | |
| Bhattacharjee et al. | Time-domain feature and ensemble model based classification of emg signals for hand gesture recognition | |
| Zhao et al. | Deep Learning with Attention on Hand Gesture Recognition Based on sEMG | |
| Hossain et al. | A comparative study of motor imagery (MI) detection in electroencephalogram (EEG) signals using different classification algorithms | |
| Narudin et al. | Brainwave Classification of Task Performed by Stroke Patients using AN | |
| Tan et al. | Classification of parkinsonian rigidity using adaboost with decision stumps |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| RJ01 | Rejection of invention patent application after publication | ||
| RJ01 | Rejection of invention patent application after publication |
Application publication date: 20191203 |