CN110507296A - A kind of acute hypotension mixing method for early warning based on LSTM network - Google Patents
A kind of acute hypotension mixing method for early warning based on LSTM network Download PDFInfo
- Publication number
- CN110507296A CN110507296A CN201910738555.9A CN201910738555A CN110507296A CN 110507296 A CN110507296 A CN 110507296A CN 201910738555 A CN201910738555 A CN 201910738555A CN 110507296 A CN110507296 A CN 110507296A
- Authority
- CN
- China
- Prior art keywords
- blood pressure
- low blood
- early warning
- data
- acute hypotension
- 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
- A61B5/02—Detecting, measuring or recording pulse, heart rate, blood pressure or blood flow; Combined pulse/heart-rate/blood pressure determination; Evaluating a cardiovascular condition not otherwise provided for, e.g. using combinations of techniques provided for in this group with electrocardiography or electroauscultation; Heart catheters for measuring blood pressure
- A61B5/021—Measuring pressure in heart or blood vessels
-
- 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
- A61B5/7264—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
-
- A—HUMAN NECESSITIES
- A61—MEDICAL OR VETERINARY SCIENCE; HYGIENE
- A61B—DIAGNOSIS; SURGERY; IDENTIFICATION
- A61B5/00—Measuring for diagnostic purposes; Identification of persons
- A61B5/74—Details of notification to user or communication with user or patient ; user input means
- A61B5/746—Alarms related to a physiological condition, e.g. details of setting alarm thresholds or avoiding false alarms
-
- A—HUMAN NECESSITIES
- A61—MEDICAL OR VETERINARY SCIENCE; HYGIENE
- A61B—DIAGNOSIS; SURGERY; IDENTIFICATION
- A61B2576/00—Medical imaging apparatus involving image processing or analysis
- A61B2576/02—Medical imaging apparatus involving image processing or analysis specially adapted for a particular organ or body part
- A61B2576/023—Medical imaging apparatus involving image processing or analysis specially adapted for a particular organ or body part for the heart
Landscapes
- Health & Medical Sciences (AREA)
- Life Sciences & Earth Sciences (AREA)
- Engineering & Computer Science (AREA)
- Physics & Mathematics (AREA)
- Molecular Biology (AREA)
- General Health & Medical Sciences (AREA)
- Biophysics (AREA)
- Pathology (AREA)
- Artificial Intelligence (AREA)
- Biomedical Technology (AREA)
- Heart & Thoracic Surgery (AREA)
- Medical Informatics (AREA)
- Veterinary Medicine (AREA)
- Surgery (AREA)
- Animal Behavior & Ethology (AREA)
- Physiology (AREA)
- Public Health (AREA)
- Signal Processing (AREA)
- Cardiology (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Psychiatry (AREA)
- Vascular Medicine (AREA)
- Evolutionary Computation (AREA)
- Fuzzy Systems (AREA)
- Mathematical Physics (AREA)
- Power Engineering (AREA)
- Measuring And Recording Apparatus For Diagnosis (AREA)
Abstract
The invention discloses a kind of acute hypotension mixing method for early warning based on LSTM network; it is first directed to the physiological data series of acquisition; hypotensive episode detection is carried out to it; if being currently at low blood pressure state; export the early warning of result low blood pressure; if currently without low blood pressure state is in; carry out acute hypotension event prediction; if acute hypotension can occur for prediction subsequent period; export result low blood pressure prediction and warning; then the detection of blood pressure protection value is carried out, if exceeding protection value beyond protection value output result blood pressure, two results is returned and carries out early warning.Acute hypotension of the present invention carries out prediction and warning, to race against time for care provider, guarantees safety of the patient in rehabilitation training, heart movement rehabilitation is made to become more scientific, efficient and safe.
Description
Technical field
The present invention relates to field of medical technology more particularly to a kind of acute hypotension mixing early warning based on LSTM network
Method.
Background technique
Blood pressure is the important indicator of a personal health condition, is also one of the main feature for reflecting cardiovascular disease.Blood
It presses through height and is easy to happen the emergency situations such as myocardial infarction, heart failure, cerebral hemorrhage.It is brought to prevent the disadvantageous changes of blood pressure
Serious consequence is in advance predicted blood pressure extremely important to monitor blood pressure.
With the progress for the treatment of cardiovascular disease theory, more and more medical workers start to advocate cardiac rehabilitation.The heart
Dirty athletic rehabilitation is also increasingly valued by people as one of big processing method of cardiac rehabilitation five.Heart movement rehabilitation pair
Quality of life of patients is improved, helps patient to return to normal life and plays an important role.Scientific and effective rehabilitation exercise, can
The aerobic exercise endurance for improving cardiac, boosts metabolism, and improves prognosis quality, reduces the death rate.In heart movement
In rehabilitation, acute hypotension is the common complication of heart disease, is common in the patient that constitution is weaker or is in a bad way.Low blood pressure
It is easy to cause dizziness, limb is soft, falls down, and even results in heart ischemia and triggers the generation of heart abnormality event.
For this purpose, there is the research predicted by blood-pressure measurement data pressure value, this aspect research in the prior art
Including the model prediction based on machine learning algorithm and neural network algorithm progress.Wherein, pass through returning in machine learning algorithm
Sound state network carries out blood pressure prediction, which is to solve conventional recursive neural network (RNN) and gradient occur
Disappear with gradient explosion issues and propose, reserve pool is added in RNN to remember to simple time series data, but this
Network can only carry out short-term memory, cannot handle complicated dynamic problem very well；The BP mind in neural network algorithm is respectively adopted
Through network and radial base neural net, the diastolic pressure situation of user is predicted by using userspersonal information, to higher than diastolic pressure
The user of normal range (NR) reminds, and to user information and the opening relationships of diastolic pressure situation, but this method only has single prediction
Ability, blood pressure early warning cannot be provided in time for user by not having prediction, the ability of real time monitoring in real time.
In conclusion to solve the above problems, and guarantee that patient can carry out safe and effective rehabilitation exercise training, this
A kind of field urgently acute hypotension method for early warning in heart movement rehabilitation,
Summary of the invention
The object of the present invention is to provide a kind of acute hypotension mixing method for early warning based on LSTM network, to urgency
Property low blood pressure carry out prediction and warning, to race against time for care provider, guarantee safety of the patient in rehabilitation training, make the heart
Dirty athletic rehabilitation becomes more scientific, efficient and safe.
The technical solution adopted by the present invention to solve the technical problems is: a kind of acute hypotension based on LSTM network
Mix method for early warning, the acute hypotension mixing method for early warning the following steps are included:
Step A imports the physiological data series of acquisition, carries out hypotensive episode detection to it, is currently at if detecting
Low blood pressure state, then exporting result is low blood pressure early warning, enters step C, otherwise, enters step B；
Step B carries out acute hypotension event prediction to current physiological data series collected, if predicting next
Acute hypotension can occur for the period, then exporting result is low blood pressure prediction and warning, then enter step C；
Step C carries out the detection of blood pressure protection value, if exceeding protection value, exports result blood pressure beyond protection value, returns
Two results in step A and step B carry out early warning, which is low blood pressure threshold value；
Wherein, predict that process includes Early-warning Model being established based on LSTM network, and carry out to the model in the step B
Training, which successively includes for step X1, step X2, step X3, step X4；
Step X1 imports the physiological data series of acquisition, pre-processes to the data；
Step X2 carries out wavelet decomposition to pretreated signal；
Step X3 is predicted by LSTM neural network structure to the detail coefficients of lower a period of time and close to coefficient；
Step X4, the data predicted previous step carry out wavelet reconstruction, and then obtain signal time sequence prediction knot
Fruit.
Preferably, step A further includes the detection process packet to whether being in acute hypotension state and detect
Include following steps:
Step 1, one is fixed in the queue of the physiological data series according to the physiological data series for importing acquisition
For pointer as starting point, tail of the queue uses integer variable N start of record into the mean arterial pressure sequence of terminal as terminal
Pressure value is lower than the data amount check of low blood pressure threshold value；
Step 2, when there are data to join the team every time, it is low to check whether the mean arterial pressure of the data of fixed pointers meaning is lower than
Blood pressure thresholds, if so, integer variable N subtracts 1, otherwise, without plus-minus, new data of joining the team；
Step 3, if new pressure value is less than low blood pressure threshold value, then variable integer variable N adds 1, otherwise without plus-minus；
Step 4, judge whether integer variable N and the ratio in reaction time are greater than low blood pressure threshold percentage, if so,
Output result is low blood pressure state.
Preferably, step A further includes carrying out prediction and warning to acute hypotension event, which is that will be imported
The physiological data series of acquisition are put into the middle of the model that step X is trained and are predicted comprising following process:
For the physiological signal data of patient in a period of time, judge whether its mean arterial pressure meetsThis is average
The calculation formula of angiosthenia are as follows:
Wherein, MAP is mean arterial pressure；Diastolic is diastolic pressure；Systolic is systolic pressure；P indicates low blood pressure
Threshold value, unit mmHg；T indicates the reaction time, and unit is minute；Q indicates low blood pressure threshold percentage.
Preferably, data are pre-processed in step X1 the following steps are included:
Step 1, the outlier is rejected there are outlier in the physiological data series for importing acquisition, meanwhile, it will put down
Equal angiosthenia is higher than 140mmHg and the data lower than 35mmHg are set to zero；
Step 2, missing values are filled, the value of missing is filled with using linear interpolation, specific formula is as follows:
Wherein, (x1, y1) and (x0, y0) indicate to lack the endpoint at section both ends.
Preferably, the decomposable process of step X2 and the restructuring procedure of X4 are as shown in figure 4, its decomposition result are as follows:
X (t)=A1(t)+D1(t)
=A2(t)+D2(t)+D1(t)
=A3(t)+D3(t)+D3(t)+D1(t)
=AL(t)+DL(t)+DL-1(t)+…+D1(t)
Wherein, H [] and L [] is the high-pass filter and low-pass filter in decomposable process, H ' [] and L ' respectively
[] is high-pass filter and low-pass filter in restructuring procedure respectively；DL (t) and AL (t) is the thin of L layers of decomposition respectively
It saves coefficient and close to coefficient, the two new coefficient sequences will be obtained and go to predict the decomposition coefficient of subsequent period as input.
It preferably, include input layer, hidden layer and output layer in LSTM network, in the nerve of hidden layer in step X3
Addition and out gate, input gate and forgetting door in member, while the parameter in LSTM network structure is adjusted, adjusting parameter
Include:
Activation layers, be tanh by its activation primitive default setting；
Recurrent_activation circulation step, is hard_sigmoid by its activation primitive default setting；
Dropout is added, the neuron in training is dropped according to probability P, P 0.5；
Timestep setting, it is associated with the input data of the time series of its last period to be set as each data；
It is log 2N~2N that range, which is arranged, in hidden layer number, and N is input layer number；
Batch_size setting, the number of samples sum being set as in this training pattern；
Epochs setting, the setting correspond to the number that all samples are completely trained；
Loss Function setting, setting Loss Function be MAE, when Loss Function restrain when, i.e.,
Stop the training of model；
Optimizer setting, using Adam.
The beneficial effects of the present invention are:
1, it by being prejudged in advance to the hypotensive episode of patient, when being abnormal situation, can give warning in advance,
The time is striven for care provider, has avoided unexpected generation, ensure that patient is capable of the carry out rehabilitation exercise of highly effective and safe；
2, low blood pressure prediction is carried out to tester by Lstm model, data are reliable, and prediction is accurate, high-efficient；
3, pass through hypotension detection early warning, the mixing alarm mode of low blood pressure prediction and warning and blood pressure protection value early warning, energy
Enough to carry out forewarning function to low blood pressure anomalous event to the greatest extent, avoiding single alarm mode can not be comprehensively to low blood
Pressure carries out early warning, and then can occasion a delay to the state of an illness.
Detailed description of the invention
In order to more clearly explain the embodiment of the invention or the technical proposal in the existing technology, below in conjunction with attached drawing and
The invention will be further described for embodiment, and the accompanying drawings in the following description is only section Example of the invention, for ability
For the those of ordinary skill of domain, without creative efforts, it is attached to can also be obtained according to these attached drawings other
Figure:
Fig. 1 is that the acute hypotension in a kind of acute hypotension mixing method for early warning based on LSTM network of the present invention is pre-
Alert flow chart；
Fig. 2 is that the present invention is in a kind of acute hypotension mixing method for early warning based on LSTM network of the present invention to low blood
Press the decision flow chart of event detection；
Fig. 3 is that the present invention is LSTM in a kind of acute hypotension mixing method for early warning based on LSTM network of the present invention
Network structure partial schematic diagram；
Fig. 4 is point that the present invention is step X2 in a kind of acute hypotension mixing method for early warning based on LSTM network of the present invention
The process schematic of the restructuring procedure of solution preocess and X4.
Specific embodiment
In order to keep the purposes, technical schemes and advantages of the embodiment of the present invention clearer, implement below in conjunction with the present invention
Technical solution in example carries out clear, complete description, it is clear that and described embodiment is section Example of the invention, and
It is not all of embodiment.Based on the embodiment of the present invention, those of ordinary skill in the art are not before making the creative labor
Every other embodiment obtained is put, protection scope of the present invention is belonged to.
In embodiment 1, a kind of acute hypotension mixing method for early warning based on LSTM network, the acute hypotension
Mix method for early warning the following steps are included:
Step A imports the physiological data series of acquisition, carries out hypotensive episode detection to it, is currently at if detecting
Low blood pressure state, then exporting result is low blood pressure early warning, enters step C, otherwise, enters step B；
Step B carries out acute hypotension event prediction to current physiological data series collected, if predicting next
Acute hypotension can occur for the period, then exporting result is low blood pressure prediction and warning, then enter step C；
Step C carries out the detection of blood pressure protection value, if exceeding protection value, exports result blood pressure beyond protection value, returns
Two results in step A and step B carry out early warning, which is low blood pressure threshold value；
Wherein, predict that process includes Early-warning Model being established based on LSTM network, and carry out to the model in the step B
Training, which successively includes for step X1, step X2, step X3, step X4；
Step X1 imports the physiological data series of acquisition, pre-processes to the data；
Step X2 carries out wavelet decomposition to pretreated signal；
Step X3 is predicted by LSTM neural network structure to the detail coefficients of lower a period of time and close to coefficient；
Step X4, the data predicted previous step carry out wavelet reconstruction, and then obtain signal time sequence prediction
As a result.
Acute hypotension mixing early warning includes hypotensive episode detection early warning, the early warning of acute hypotension event prediction and blood
Pressure value is more than the early warning of protection value.According to the definition of acute hypotension, i.e. pressure value occurs from normal or high value for some reason
It is decreased obviously, the vitals such as brain, heart, kidney make one to occur that dizziness, pupil, limb be soft, cold sweat, palpitaition, oliguresis because of ischemic
Etc. symptoms, serious person occur fainting or suffer a shock, low blood pressure state is judged by using mean arterial pressure, data are reliable, symptom
It predicts calibrated.The normal mean arterial pressure range of people is 70mmHg~105mmHg, is low blood pressure state below this range,
Therefore low blood pressure threshold range is set as 70mmHg.
Whether acute hypotension state has been in for detection patient, will have been carried out by the hypotensive episode to patient
Definition, including establish data queue's testing mechanism, i.e., detected patient is pressed into reaction time sequence per hypotensive episode together
Number is discharged into data queue, as shown in Fig. 2, fixing a pointer in the queue as starting point, tail of the queue makes as terminal
Data with pressure value in the mean arterial pressure sequence of an integer variable N start of record to terminal lower than low blood pressure threshold value
Number.When there are data to join the team every time, check whether the mean arterial pressure of the data of fixed pointers meaning is lower than low blood pressure threshold value, such as
Fruit is that then integer variable N subtracts one, otherwise without plus-minus, new data of then joining the team.If new pressure value is less than low blood pressure threshold
Value, then integer variable adds one, otherwise without plus-minus.Finally, whether big by integer variable N and the ratio in reaction time
Judge whether it has been in acute hypotension state in low blood pressure threshold percentage.
Prediction and warning is carried out for acute hypotension event, it will be by being established in step X based on the pre- of LSTM network
Alert model carries out prediction and warning.Wherein, it inputs above-mentioned data queue to be predicted into model, and acute hypotension is predicted
It is defined: including the physiological signal data by giving patient in a period of time, judging being averaged for patient after this time
Whether angiosthenia meets, i.e., whether acute hypotension event can occur.Wherein p indicates low blood pressure threshold value, and unit is mmHg, t table
Show the reaction time, unit is minute, and Q indicates low blood pressure threshold percentage.Such asIndicate that the Mean Arterial of patient is pressed in
There is 70% value to be less than or equal to 70mmHg in no less than 20 minutes.
Further, step A further includes the detection process to whether being in acute hypotension state and detect
Include the following steps:
Step 1, one is fixed in the queue of the physiological data series according to the physiological data series for importing acquisition
For pointer as starting point, tail of the queue uses integer variable N start of record into the mean arterial pressure sequence of terminal as terminal
Pressure value is lower than the data amount check of low blood pressure threshold value；
Step 2, when there are data to join the team every time, it is low to check whether the mean arterial pressure of the data of fixed pointers meaning is lower than
Blood pressure thresholds, if so, integer variable N subtracts one, otherwise, without plus-minus, new data of joining the team；
Step 3, if new pressure value is less than low blood pressure threshold value, then variable integer variable N adds one, otherwise without plus-minus；
Step 4, judge whether integer variable N and the ratio in reaction time are greater than low blood pressure threshold percentage, if so,
Output result is low blood pressure state.
Further, step A further includes carrying out prediction and warning to acute hypotension event, which is that will be led
The physiological data series for entering acquisition, which are put into the middle of the model that step X is trained, to be predicted comprising following process:
For the physiological signal data of patient in a period of time, judge whether its mean arterial pressure meetsThis is average
The calculation formula of angiosthenia are as follows:
Wherein, MAP is mean arterial pressure；Diastolic is diastolic pressure；Systolic is systolic pressure；P indicates low blood pressure
Threshold value, unit mmHg；T indicates the reaction time, and unit is minute；Q indicates low blood pressure threshold percentage.
Further, data are pre-processed in step X1 the following steps are included:
Step 1, the outlier is rejected there are outlier in the physiological data series for importing acquisition, meanwhile, it will put down
Equal angiosthenia is higher than 140mmHg and the data lower than 35mmHg are set to zero；
Step 2, missing values are filled, the value of missing is filled with using linear interpolation, specific formula is as follows:
Wherein, (x1, y1) and (x0, y0) indicate to lack the endpoint at section both ends.
Specifically, in step 1, it is contemplated that during acquiring patient's blood pressure signal, exist because external action causes
Signal errors, meanwhile, when the codomain of signal collected much aberrant blood pressure measurement, such as higher than 140mmHg or it is lower than
The pressure value of 35mmHg appears to be extremely abnormal, unreasonable on medical angle, so by such data zero setting, Yi Mianying
Ring the training of sample mileage evidence in Early-warning Model.
Further, the decomposable process of step X2 and the restructuring procedure of X4 are as shown in figure 4, its decomposition result are as follows:
X (t)=A1(t)+D1(t)
=A2(t)+D2(t)+D1(t)
=A3(t)+D3(t)+D3(t)+D1(t)
=AL(t)+DL(t)+DL-1(t)+…+D1(t)
Wherein, H [] and L [] is the high-pass filter and low-pass filter in decomposable process, H ' [] and L ' respectively
[] is high-pass filter and low-pass filter in restructuring procedure respectively；DL (t) and AL (t) is the thin of L layers of decomposition respectively
It saves coefficient and close to coefficient, the two new coefficient sequences will be obtained and go to predict the decomposition coefficient of subsequent period as input.
Specifically, in catabolic phase, low-pass filter removes the high frequency section output low frequency part of input signal, high pass
Filter filters low frequency part output high frequency section, then that the progress of filtered signal is down-sampled twice, obtains approximation coefficient
And detail coefficients.Reconstruct is the inverse process decomposed, wherein for the filter of high quality, X (t)=X ' (t).
It further, include input layer, hidden layer and output layer in LSTM network, in the mind of hidden layer in step X3
Through addition in member and out gate, input gate and forgetting door, while the parameter in LSTM network structure is adjusted, adjustment ginseng
Number includes:
Activation layers, be tanh by its activation primitive default setting；
Recurrent_activation circulation step, is hard_sigmoid by its activation primitive default setting；
Dropout is added, the neuron in training is dropped according to probability P, P 0.5；
Timestep setting, it is associated with the input data of the time series of its last period to be set as each data；
It is log 2N~2N that range, which is arranged, in hidden layer number, and N is input layer number；
Batch_size setting, the number of samples sum being set as in this training pattern；
Epochs setting, the setting correspond to the number that all samples are completely trained；
Loss Function setting, setting Loss Function be MAE, when Loss Function restrain when, i.e.,
Stop the training of model；
Optimizer setting, using Adam；
Specifically, by Lstm neural network structure go to predict the detail coefficients of following a period of time and close to coefficient.
LSTM is one deformation of RNN.RNN is the most effective tool for handling time series related data, compared to other nerves
Network, the result of the output layer of RNN is not only and current input is related but also related with last hidden layer result, thus phase
When in there is certain memory function to time series.By the way that three valves are arranged in LSTM network, to act on RNN's
Whether the network memory state before adjusting on node acts on the calculating of current network.As shown in figure 3, small circle indicates
The valve of addition.Activation layers are active coating, setting of this layer for activation primitive in LSTM network, unbalanced input
Function, to complete Nonlinear Mapping, wherein tanh is hyperbolic tangent function, is set for the default activation function of this active coating
It sets；Recurrent_activation is the activation primitive that circulation step applies, and hard_ is arranged using default function
Sigmoid, the function are the piece wire approximations in logic activation function, the pace of learning for improving model are acted on, with police
It wakes up faster acute hypotension prediction and warning；In view of if the parameter of model is too many, and training in the training of model
Sample is again very little, trains the model come and is easy to the phenomenon that generating over-fitting, and then leads to loss function ratio in test data
Larger, predictablity rate is lower, while in order to improve accuracy rate, the instruction to require a great deal of time to sample progress repeatedly
Practice, so here, using Dropout, specifically including forward direction in a model to reduce over-fitting and time loss problem
In communication process, the activation value of some neuron is allowed to stop working with certain Probability p, and then prevent it from depending on part unduly
Feature, and then keep model generalization stronger；Timestep is the number of the input of each data and how many before time serieses
According to being associated, hidden layer number is depended between reference value 2N and log 2N, and N is input layer number, works as node in hidden layer
When setting is less than log 2N, the fitting effect of network can decline；When setting is greater than 2N, the training time can extend, and easily fall into
Local minimum point；Batch_size indicates once trained number of samples, which influences whether the degree of optimization and instruction of model
Practice speed, Batch_size value is bigger, and training speed can be faster, i.e. resultant error convergence is faster, but the extensive energy of model
Power can reduce, so the value is depending on data volume when detection；Epochs is the number of iterations, that is, refers to and use all samples
Completely trained number, setting Loss Function (loss function) is MAE (mean absolute error), to indicate training result
Error stops instruction when Loss function is in convergence；Optimizer is optimizer, which is set as Adam, based on
It calculates and updates step-length, compared to optimizers such as MomentumAdagrad, Adadelta, RMSprop, which raises model learning effects
Rate strengthens adaptability.In addition, calculate prediction data and truthful data root-mean-square error (RMSE) Lai Hengliang predicted value and
Deviation between actual value, RMSE are the root-mean-square error of prediction data and truthful data, which is used to evaluate the variation of data
Degree, RMSE is smaller, illustrate that prediction model prediction result precision is higher, adjusts the parameter in LSTM network with this.
Claims (6)
1. a kind of acute hypotension mixing method for early warning based on LSTM network, which is characterized in that the acute hypotension mixing
Method for early warning the following steps are included:
Step A imports the physiological data series of acquisition, carries out hypotensive episode detection to it, is currently at low blood if detecting
Pressure condition, then exporting result is low blood pressure early warning, enters step C, otherwise, enters step B；
Step B carries out acute hypotension event prediction to current physiological data series collected, if predicting subsequent period meeting
Acute hypotension occurs, then exporting result is low blood pressure prediction and warning, then enters step C；
Step C carries out the detection of blood pressure protection value, if exceeding protection value, exports result blood pressure beyond protection value, return step A
Early warning is carried out with two results in step B, which is low blood pressure threshold value；
Wherein, predict that process includes establishing Early-warning Model based on LSTM network, and be trained to the model in the step B,
The training step successively includes for step X1, step X2, step X3, step X4；
Step X1 imports the physiological data series of acquisition, pre-processes to the data；
Step X2 carries out wavelet decomposition to pretreated signal；
Step X3 is predicted by LSTM neural network structure to the detail coefficients of lower a period of time and close to coefficient；
Step X4, the data predicted previous step carry out wavelet reconstruction, and then obtain signal time sequence prediction result.
2. the acute hypotension mixing method for early warning according to claim 1 based on LSTM network, which is characterized in that step
Whether A further includes to being in acute hypotension state and detect, which includes the following steps:
Step 1, a pointer is fixed in the queue of the physiological data series according to the physiological data series for importing acquisition
As starting point, tail of the queue uses integer variable N start of record pressure value into the mean arterial pressure sequence of terminal as terminal
Lower than the data amount check of low blood pressure threshold value；
Step 2, when there are data to join the team every time, check whether the mean arterial pressure of the data of fixed pointers meaning is lower than low blood pressure threshold
Value, if so, integer variable N subtracts one, otherwise, without plus-minus, new data of joining the team；
Step 3, if new pressure value is less than low blood pressure threshold value, then variable integer variable N adds one, otherwise without plus-minus；
Step 4, judge whether integer variable N and the ratio in reaction time are greater than low blood pressure threshold percentage, if so, output knot
Fruit is low blood pressure state.
3. the acute hypotension mixing method for early warning according to claim 1 or 2 based on LSTM network, which is characterized in that
Step A further includes that prediction and warning is carried out to acute hypotension event, which is that will import the physiological data sequence acquired
Column, which are put into the middle of the model that step X is trained, to be predicted comprising following process:
For the physiological signal data of patient in a period of time, judge whether its mean arterial pressure meetsThe mean arterial pressure
Calculation formula are as follows:
Wherein, MAP is mean arterial pressure；Diastolic is diastolic pressure；Systolic is systolic pressure；P indicates low blood pressure threshold value
, unit mmHg；T indicates the reaction time, and unit is minute；Q indicates low blood pressure threshold percentage.
4. the acute hypotension mixing method for early warning according to claim 1 based on LSTM network, which is characterized in that step
Data are pre-processed in X1 the following steps are included:
Step 1, the outlier is rejected there are outlier in the physiological data series for importing acquisition, meanwhile, by Mean Arterial
Pressure is higher than 140mmHg and the data lower than 35mmHg are set to zero；
Step 2, missing values are filled, the value of missing is filled with using linear interpolation, specific formula is as follows:
Wherein, (x1, y1) and (x0, y0) indicate to lack the endpoint at section both ends.
5. the acute hypotension mixing method for early warning according to claim 1 based on LSTM network, which is characterized in that step
The decomposable process of X2 and the restructuring procedure of X4 are as shown in figure 4, its decomposition result are as follows:
X (t)=A1(t)+D1(t)
=A2(t)+D2(t)+D1(t)
=A3(t)+D3(t)+D3(t)+D1(t)
=AL(t)+DL(t)+DL-1(t)+...+D1(t)
Wherein, H [] and L [] is the high-pass filter and low-pass filter in decomposable process, H ' [] and L ' [] respectively
It is the high-pass filter and low-pass filter in restructuring procedure respectively；DL (t) and AL (t) is the details system of L layers of decomposition respectively
It counts and close to coefficient, the decomposition coefficient that the two new coefficient sequences remove prediction subsequent period as input will be obtained.
6. the acute hypotension mixing method for early warning according to claim 1 based on LSTM network, which is characterized in that in step
Include input layer, hidden layer and output layer in LSTM network in rapid X3, is added on the neuron of hidden layer and out gate, defeated
Introduction and forgetting door, while the parameter in LSTM network structure is adjusted, adjusting parameter includes:
Activation layers, be tanh by its activation primitive default setting；
Recurrent_activation circulation step, is hard_sigmoid by its activation primitive default setting；
Dropout is added, the neuron in training is dropped according to probability P, P 0.5；
Timestep setting, it is associated with the input data of the time series of its last period to be set as each data；
It is log 2N~2N that range, which is arranged, in hidden layer number, and N is input layer number；
Batch_size setting, the number of samples sum being set as in this training pattern；
Epochs setting, the setting correspond to the number that all samples are completely trained；
Loss Function setting, setting Loss Function be MAE, when Loss Function convergence when, that is, stop mould
The training of type；
Optimizer setting, using Adam.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201910738555.9A CN110507296A (en) | 2019-08-12 | 2019-08-12 | A kind of acute hypotension mixing method for early warning based on LSTM network |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201910738555.9A CN110507296A (en) | 2019-08-12 | 2019-08-12 | A kind of acute hypotension mixing method for early warning based on LSTM network |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN110507296A true CN110507296A (en) | 2019-11-29 |
Family
ID=68624968
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201910738555.9A Pending CN110507296A (en) | 2019-08-12 | 2019-08-12 | A kind of acute hypotension mixing method for early warning based on LSTM network |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN110507296A (en) |
Cited By (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN111728602A (en) * | 2020-08-21 | 2020-10-02 | 之江实验室 | Non-contact blood pressure measuring device based on PPG |
| CN112037918A (en) * | 2020-11-02 | 2020-12-04 | 平安科技（深圳）有限公司 | Chronic disease medical insurance cost prediction method fusing complication risks and related equipment |
| CN114937486A (en) * | 2022-06-22 | 2022-08-23 | 肾泰网健康科技(南京)有限公司 | Construction method and application of IDH prediction and intervention measure recommendation multitask model |
| WO2023015372A1 (en) * | 2021-08-12 | 2023-02-16 | Ottawa Heart Institute Research Corporation | Systems, methods and apparatus for predicting hemodynamic events |
| WO2024032110A1 (en) * | 2022-08-12 | 2024-02-15 | 宁波越凡医疗科技有限公司 | Method and device for controlling blood pressure |
-
2019
- 2019-08-12 CN CN201910738555.9A patent/CN110507296A/en active Pending
Cited By (6)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN111728602A (en) * | 2020-08-21 | 2020-10-02 | 之江实验室 | Non-contact blood pressure measuring device based on PPG |
| CN112037918A (en) * | 2020-11-02 | 2020-12-04 | 平安科技（深圳）有限公司 | Chronic disease medical insurance cost prediction method fusing complication risks and related equipment |
| WO2023015372A1 (en) * | 2021-08-12 | 2023-02-16 | Ottawa Heart Institute Research Corporation | Systems, methods and apparatus for predicting hemodynamic events |
| CN114937486A (en) * | 2022-06-22 | 2022-08-23 | 肾泰网健康科技(南京)有限公司 | Construction method and application of IDH prediction and intervention measure recommendation multitask model |
| CN114937486B (en) * | 2022-06-22 | 2023-09-26 | 肾泰网健康科技(南京)有限公司 | IDH prediction and intervention measure recommendation multi-task model construction method and application |
| WO2024032110A1 (en) * | 2022-08-12 | 2024-02-15 | 宁波越凡医疗科技有限公司 | Method and device for controlling blood pressure |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN110507296A (en) | A kind of acute hypotension mixing method for early warning based on LSTM network | |
| Zong et al. | Reduction of false arterial blood pressure alarms using signal quality assessement and relationships between the electrocardiogram and arterial blood pressure | |
| Němcová et al. | A comparative analysis of methods for evaluation of ECG signal quality after compression | |
| Struzik et al. | 1∕ f scaling in heart rate requires antagonistic autonomic control | |
| CN108478209A (en) | Ecg information dynamic monitor method and dynamic monitor system | |
| CN108577830A (en) | A kind of user oriented sign information dynamic monitor method and dynamic monitor system | |
| CN114081462B (en) | Heart health monitoring system based on multidimensional physiological information | |
| Solem et al. | Prediction of intradialytic hypotension using photoplethysmography | |
| Chen et al. | A machine learning method correlating pulse pressure wave data with pregnancy | |
| CN104434312A (en) | Monitoring equipment and physiological parameter processing method and system thereof | |
| Liu et al. | Cuffless blood pressure measurement using smartwatches: a large-scale validation study | |
| CN112656393A (en) | Heart rate variability detection method and system | |
| Baek et al. | Blood pressure prediction by a smartphone sensor using fully convolutional networks | |
| Kenneth et al. | Data fusion of multimodal cardiovascular signals | |
| CN113040738B (en) | Blood pressure detecting device | |
| Suzuki et al. | Role of vagal control in vasovagal syncope | |
| Pytel et al. | Anthropometric predictors and artificial neural networks in the diagnosis of hypertension | |
| CN115836847A (en) | Blood pressure prediction device and equipment | |
| Nollo et al. | Synchronization index for quantifying nonlinear causal coupling between RR interval and systolic arterial pressure after myocardial infarction | |
| Salinger et al. | Microcomputer system for telemetric assessment of short term heart rate variability in time and frequency domain, type VariaCardio TF4 | |
| Han et al. | Noncontact blood pressure estimation using BP-related cardiovascular knowledge: An uncalibrated method based on consumer-level camera | |
| CN111710419A (en) | Heart health quality early warning system and early warning method thereof | |
| CN110547779A (en) | Trip health detection system based on Internet of things | |
| CN114831643B (en) | Electrocardiosignal monitoring devices and wearable equipment | |
| WO2018236252A1 (en) | Method of assessing the risk of complications of diseases of the cardiovascular system with concomitant pathology |
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 |