An acute hypotension mixed early warning method based on an LSTM network

Abstract

The application discloses an acute hypotension mixed early warning method based on an LSTM network, first, the collected physiological data sequence is introduced, and hypotension event detection is carried out on the physiological data sequence, if the current is in a hypotension state, a low blood pressure early warning result is output, if the current is not in the hypotension state, acute hypotension event prediction is carried out, if it is predicted that acute hypotension will occur in the next period, a low blood pressure prediction early warning result is output, then blood pressure protection value detection is carried out, if the blood pressure exceeds the protection value, a blood pressure exceeding protection value result is output, and the two results are returned to carry out early warning. The application carries out prediction and early warning on acute hypotension, thereby striving for time for the monitor, guaranteeing the safety of the patient during rehabilitation training, and making heart motion rehabilitation more scientific, efficient and safe.

Description

TECHNICAL FIELD

[0001] The application relates to the technical field of medical treatment, in particular to an acute hypotension mixed early warning method based on an LSTM network. BACKGROUND

[0002] With the progress of the treatment concept of cardiovascular diseases, more and more medical workers begin to advocate heart rehabilitation. As one of the five processing methods of heart rehabilitation, heart exercise rehabilitation is also paid more and more attention by people. Heart exercise rehabilitation plays an important role in improving the quality of life of patients and helping patients return to normal life. Scientific and effective rehabilitation exercise can improve the aerobic exercise endurance of heart disease patients, promote metabolism, improve prognosis quality and reduce mortality. In heart exercise rehabilitation, acute hypotension is a common complication of heart disease, which is common in patients with weak constitution or serious illness. Hypotension is easy to cause dizziness, flaccid limbs, fall, and even trigger heart ischemia and abnormal events.

[0003] Therefore, the prior art has appeared researches on predicting blood pressure values through blood pressure measurement data, which include model prediction based on machine learning algorithms and neural network algorithms. Among them, blood pressure prediction is performed through an echo state network in the machine learning algorithm. The echo state network is proposed to solve the gradient disappearance and gradient explosion problems of traditional recurrent neural networks (RNN), and a reservoir is added in the RNN to memorize simple time series data, but the network can only perform short-term memory and cannot well handle complex dynamic problems. The BP neural network and the radial basis neural network in the neural network algorithm are used respectively, the diastolic pressure of the user is predicted by using the personal information of the user, the user who is higher than the normal range of diastolic pressure is reminded, and the relationship between the user information and the diastolic pressure is established, but the method only has single prediction capability and does not have real-time prediction and real-time monitoring capability, and cannot provide blood pressure early warning for the user in time.

[0004] In view of the above, in order to solve the above problems and ensure that patients can perform safe and effective rehabilitation exercise training, the technical field urgently needs an acute hypotension mixed early warning method based on an LSTM network, SUMMARY

[0005] The application aims to provide an acute hypotension mixed early warning method based on an LSTM network, which can predict and warn acute hypotension, thereby gaining time for the guardian, ensuring the safety of patients during rehabilitation training, and making heart exercise rehabilitation more scientific, efficient and safe.

[0006] The technical scheme adopted by the application to solve the technical problems is: an acute hypotension mixed early warning method based on an LSTM network, the acute hypotension mixed early warning method comprising the following steps:

[0007] Step A, import the collected physiological data sequence, detect the hypotension event, if the current is in hypotension state, output the result as hypotension warning, enter step C, otherwise, enter step B;

[0008] Step B, acute hypotension event prediction is performed on the current collected physiological data sequence, if the next period is predicted to occur acute hypotension, the output result is hypotension prediction warning, and then enter step C;

[0009] Step C, blood pressure protection value detection, if it exceeds the protection value, the output result is blood pressure exceeding the protection value, return to step A and step B for warning, the blood pressure protection value is the hypotension threshold;

[0010] Wherein, the prediction process in step B includes establishing a warning model based on LSTM network, and training the model, the training step includes step X1, step X2, step X3, step X4 in turn;

[0011] Step X1, import the collected physiological data sequence, and pre-process the data;

[0012] Step X2, wavelet decomposition is performed on the pre-processed signal;

[0013] Step X3, the next time detail coefficient and proximity coefficient are predicted by LSTM neural network structure;

[0014] Step X4, wavelet reconstruction is performed on the data predicted in the previous step, and then the signal time sequence prediction result is obtained.

[0015] Preferably, step A further includes detecting whether it has been in acute hypotension state, the detection process includes the following steps:

[0016] Step 1, according to the imported collected physiological data sequence, fix a pointer as the starting point and the tail as the end point in the queue of the physiological data sequence, and use an integer variable N to record the number of data whose blood pressure value is lower than the hypotension threshold in the mean arterial pressure sequence from the starting point to the end point;

[0017] Step 2, every time data is enqueued, check whether the mean arterial pressure of the data pointed by the fixed pointer is lower than the hypotension threshold, if yes, the integer variable N is reduced by 1, otherwise, no addition or subtraction is performed, and the new data is enqueued;

[0018] Step 3, if the new blood pressure value is less than the hypotension threshold, then the integer variable N is increased by 1, otherwise no addition or subtraction is performed;

[0019] Step 4, judge whether the ratio of integer variable N and reaction time is greater than the hypotension threshold percentage, if yes, the output result is low blood pressure state.

[0020] Preferably, step A further comprises a predictive warning of acute hypotension event, which is to put the imported collected physiological data sequence into the model trained in step X to make a prediction, including the following process:

[0021] For a period of time of the patient's physiological signal data, determine whether the mean arterial pressure satisfies The formula for calculating the mean arterial pressure is: ;

[0022] Wherein, MAP is the mean arterial pressure; Diastolic is diastolic blood pressure; Systolic is systolic blood pressure; p represents the hypotension threshold, with the unit of mmHg; t represents the reaction time, with the unit of minutes; q represents the percentage of hypotension threshold.

[0023] Preferably, the data preprocessing in step X1 includes the following steps:

[0024] Step 1: In the imported collected physiological data sequence, there are outliers, which are removed, and the data with mean arterial pressure higher than 140mmHg and lower than 35mmHg are set to zero;

[0025] Step 2: Fill in the missing values, and use linear interpolation to fill in the missing values, the specific formula is as follows:

[0026] ;

[0027] Wherein, (x1, y1) and (x0, y0) represent the end points at both ends of the missing segment.

[0028] Preferably, the decomposition process of step X2 and the reconstruction process of X4 are as shown in Figure 4 , and the decomposition result is:

[0029] X(t)=A1(t)+D1(t)

[0030] =A2(t)+D2(t)+D1(t)

[0031] =A3(t)+D3(t)+D3(t)+D1(t)

[0032] =A L (t)+D L (t)+D L-1 (t)+…+D1(t)

[0033] Wherein, H[·] and L[·] are high-pass filter and low-pass filter in the decomposition process respectively, H'[·] and L'[·] are high-pass filter and low-pass filter in the reconstruction process respectively; DL(t) and AL(t) are the detail coefficient and approximation coefficient of the Lth layer decomposition respectively, the two new coefficient sequences are taken as input to predict the decomposition coefficient of the next period.

[0034] Preferably, in step X3, the LSTM network comprises an input layer, a hidden layer and an output layer, an output gate, an input gate and a forget gate are added to the neurons of the hidden layer, and the parameters in the LSTM network structure are adjusted, and the adjustment parameters include:

[0035] Activation layer, the activation function of which is set as tanh by default;

[0036] Recurrent_activation loop step, the activation function of which is set as hard_sigmoid by default;

[0037] Dropout is added, and neurons in training are discarded according to a probability P, P is 0.5;

[0038] Timestep setting, which is set to be associated with the input data of each data and its previous time sequence;

[0039] The number of hidden layers is set to be in the range of log2N~2N, N is the number of input layer nodes;

[0040] Batch_size setting, which is set to the total number of sample numbers in the current training model;

[0041] Epochs setting, which corresponds to the number of complete training of all samples;

[0042] LossFunction setting, setting LossFunction as MAE, when LossFunction converges, the training of the model is stopped;

[0043] Optimizer setting, Adam is adopted.

[0044] The beneficial effects of the present application are:

[0045] 1. By predicting the hypotension event of the patient in advance, when an abnormal situation occurs, early warning can be given, time is gained for the guardian, and accidents are avoided, so that the patient can efficiently and safely recover from the movement;

[0046] 2. The Lstm model is used for hypotension prediction of the detector, the data is reliable, the prediction is accurate, and the efficiency is high;

[0047] 3. The mixed warning method of acute hypotension warning based on the LSTM network can maximize the warning effect of abnormal events of hypotension, avoids the single warning method which cannot comprehensively warn hypotension, and thus delays the disease. BRIEF DESCRIPTION OF DRAWINGS

[0048] In order to more clearly illustrate the technical solutions of the embodiments of the present application or the prior art, the present application will be further described below with reference to the drawings and embodiments. The drawings in the following description are only some embodiments of the present application, and those skilled in the art can obtain other drawings according to these drawings without creative labor.

[0049] Figure 1 FIG. 1 is a flowchart of the acute hypotension warning process in the acute hypotension mixed warning method based on the LSTM network of the present application;

[0050] Figure 2 FIG. 2 is a flowchart of the judgment of the low blood pressure event detection in the acute hypotension mixed warning method based on the LSTM network of the present application;

[0051] Figure 3 FIG. 3 is a local schematic diagram of the LSTM network structure in the acute hypotension mixed warning method based on the LSTM network of the present application;

[0052] Figure 4 FIG. 4 is a process diagram of the decomposition process of step X2 and the reconstruction process of step X4 in the acute hypotension mixed warning method based on the LSTM network of the present application. DETAILED DESCRIPTION

[0053] In order to make the purpose, technical solutions and advantages of the embodiments of the present application more clear, the technical solutions in the embodiments of the present application will be described below clearly and completely. Obviously, the described embodiments are some embodiments of the present application, but not all embodiments. Based on the embodiments of the present application, all other embodiments obtained by those skilled in the art without creative labor are within the protection scope of the present application.

[0054] In embodiment 1, the acute hypotension mixed warning method based on the LSTM network includes the following steps:

[0055] Step A: import the collected physiological data sequence, detect the low blood pressure event, if the current is in the low blood pressure state, output the result as the low blood pressure warning, and enter step C, otherwise, enter step B;

[0056] Step B, acute hypotension event prediction is performed on the current collected physiological data sequence, if it is predicted that acute hypotension will occur in the next period, the output result is a hypotension prediction warning, and then step C is entered;

[0057] Step C, blood pressure protection value detection is performed, if the protection value is exceeded, the output result is blood pressure exceeding the protection value, and the two results in steps A and B are returned for warning, the blood pressure protection value is a hypotension threshold value;

[0058] In the prediction process in step B, an early warning model is established based on an LSTM network, and the model is trained, the training steps include steps X1, X2, X3 and X4 in sequence.

[0059] Step X1, the collected physiological data sequence is imported, and the data is preprocessed;

[0060] Step X2, wavelet decomposition is performed on the preprocessed signal;

[0061] Step X3, the next time detail coefficient and proximity coefficient are predicted through an LSTM neural network structure;

[0062] Step X4, the data predicted in the previous step is wavelet reconstructed to obtain a signal time sequence prediction result.

[0063] The acute hypotension mixed warning includes hypotension event detection warning, acute hypotension event prediction warning and blood pressure value exceeding protection value warning. According to the definition of acute hypotension, i.e. the blood pressure value decreases significantly from normal or higher value due to some reasons, the brain, heart, kidney and other important organs are ischemic, and the person appears symptoms such as dizziness, black eyes, soft limbs, cold sweat, palpitation, oliguria, etc., and severe cases appear syncope or shock, the mean arterial pressure is used to judge the hypotension state, the data is reliable, and the symptom prediction is more accurate. The normal mean arterial pressure range of human is 70mmHg-105mmHg, below this range is the hypotension state, so the hypotension threshold value range is set to 70mmHg.

[0064] For detecting whether the patient has been in an acute hypotension state, the patient's hypotension event is defined, including establishing a data queue detection mechanism, i.e. each hypotension event of the detected patient is numbered and arranged in the data queue in the order of reaction time, such as Figure 2As shown, a pointer is fixed in the queue as the starting point, the end of the queue as the end point, and an integer variable N is used to record the number of data in the sequence of mean arterial pressure from the starting point to the end point whose blood pressure value is lower than the low blood pressure threshold. Each time data is enqueued, check if the mean arterial pressure of the data pointed to by the fixed pointer is lower than the low blood pressure threshold. If yes, the integer variable N is decremented by one, otherwise, no addition or subtraction is performed, and then the new data is enqueued. If the new blood pressure value is less than the low blood pressure threshold, then the integer variable is incremented by one, otherwise, no addition or subtraction is performed. Finally, whether the ratio of the integer variable N to the reaction time is greater than the low blood pressure threshold percentage is determined to determine whether it has been in an acute hypotension state.

[0065] For the prediction and early warning of acute hypotension events, the prediction and early warning model based on the LSTM network established in step X is used for prediction and early warning. Among them, the above data queue is input into the model for prediction, and the prediction of acute hypotension is defined: including judging whether the mean arterial pressure of the patient after a given period of time will satisfy, that is, whether an acute hypotension event will occur, through the physiological signal data of the patient in the given period of time. Among them, p represents the low blood pressure threshold, the unit is mmHg, t represents the reaction time, the unit is minute, and Q represents the low blood pressure threshold percentage. For example It means that the mean arterial pressure of the patient has 70% values less than or equal to 70 mmHg within not less than 20 minutes.

[0066] Further, step A further comprises detecting whether an acute hypotension state has been reached, which detection process comprises the following steps:

[0067] Step 1, according to the imported physiological data sequence, in the queue of the physiological data sequence, a pointer is fixed as the starting point, the end of the queue as the end point, and an integer variable N is used to record the number of data in the sequence of mean arterial pressure from the starting point to the end point whose blood pressure value is lower than the low blood pressure threshold;

[0068] Step 2, each time data is enqueued, check if the mean arterial pressure of the data pointed to by the fixed pointer is lower than the low blood pressure threshold. If yes, the integer variable N is decremented by one, otherwise, no addition or subtraction is performed, and then the new data is enqueued.

[0069] Step 3, if the new blood pressure value is less than the low blood pressure threshold, then the integer variable N is incremented by one, otherwise, no addition or subtraction is performed.

[0070] Step 4, whether the ratio of the integer variable N to the reaction time is greater than the low blood pressure threshold percentage is determined, and if yes, the result is output as a low blood pressure state.

[0071] Further, step A further comprises predicting and early warning of acute hypotension events, which prediction and early warning is to put the imported physiological data sequence into the model trained in step X for prediction, which comprises the following process:

[0072] For a period of time, the physiological signal data of a patient is judged whether the mean arterial pressure meets The formula for calculating the mean arterial pressure is: ;

[0073] Wherein, MAP is the mean arterial pressure; Diastolic is diastolic pressure; Systolic is systolic pressure; p represents the low blood pressure threshold, its unit is mmHg; t represents the reaction time, its unit is minute; Q represents the low blood pressure threshold percentage.

[0074] Further, the data preprocessing in step X1 includes the following steps:

[0075] Step 1, in the imported collected physiological data sequence, there are outliers, the outliers are removed, and the data with mean arterial pressure higher than 140mmHg and lower than 35mmHg is set to zero;

[0076] Step 2, fill in the missing values, for the missing values, use linear interpolation to fill in, the specific formula is as follows:

[0077] ;

[0078] Wherein, (x1, y1) and (x0, y0) represent the end points at both ends of the missing segment.

[0079] Specifically, in step 1, considering that in the process of collecting the blood pressure signal of the patient, there is signal error caused by external influence, at the same time, the collected signal is far away from the value range of conventional blood pressure measurement, for example, the blood pressure value higher than 140mmHg or lower than 35mmHg is extremely abnormal and unreasonable in medical angle, so such data is set to zero, so as to avoid affecting the training of data in the sample in the early warning model.

[0080] Further, the decomposition process of step X2 and the reconstruction process of X4 are as shown in Figure 4 The decomposition result is:

[0081] X(t)=A1(t)+D1(t)

[0082] =A2(t)+D2(t)+D1(t)

[0083] =A3(t)+D3(t)+D3(t)+D1(t)

[0084] =A L (t)+D L (t)+D L-1 (t)+…+D1(t)

[0085] Wherein, H[·] and L[·] are high-pass filter and low-pass filter in the decomposition process, H'[·] and L'[·] are high-pass filter and low-pass filter in the reconstruction process respectively; DL(t) and AL(t) are the detail coefficient and approximation coefficient of the Lth layer decomposition, and the two new coefficient sequences are taken as input to predict the decomposition coefficient of the next period.

[0086] Specifically, in the decomposition stage, the low-pass filter removes the high-frequency part of the input signal and outputs the low-frequency part, the high-pass filter filters out the low-frequency part and outputs the high-frequency part, and then the filtered signal is down-sampled twice to obtain the approximation coefficient and the detail coefficient. Reconstruction is the inverse process of decomposition, wherein, for high-quality filter, X(t)=X'(t).

[0087] Further, in step X3, the LSTM network includes an input layer, a hidden layer and an output layer, an output gate, an input gate and a forget gate are added to the neurons of the hidden layer, and the parameters in the LSTM network structure are adjusted, the adjustment parameters include:

[0088] Activation layer, the activation function of which is set to tanh by default;

[0089] Recurrent_activation loop step, the activation function of which is set to hard_sigmoid by default;

[0090] Add Dropout, neurons in training are discarded according to the probability P, P is 0.5;

[0091] Timestep setting, set to associate each data with the input data of its previous time sequence;

[0092] The number of hidden layers is set to the range of log2N~2N, N is the number of input layer nodes;

[0093] Batch_size setting, set to the total number of sample number in this training model;

[0094] Epochs setting, which corresponds to the number of complete training of all samples;

[0095] LossFunction setting, set LossFunction to MAE, when LossFunction converges, stop training the model;

[0096] Optimizer setting, Adam is adopted;

[0097] Specifically, the detail coefficient and the proximity coefficient in the future period are predicted by the Lstm neural network structure. The LSTM is a variant of RNN. The RNN is the most effective tool for processing time series related data. Compared with other neural networks, the output layer of the RNN is not only related to the current input but also related to the last hidden layer result, which is equivalent to having a certain memory function for time series. By setting three valves in the LSTM network, the previous network memory state is adjusted to act on the current network calculation. Figure 3The small circles represent the added valves. The Activation layer is the activation layer, which is used to set the activation function in the LSTM network, input a nonlinear function to complete the nonlinear mapping, where tanh is the hyperbolic tangent function, which is the default activation function setting of this activation layer; Recurrent_activation is the activation function applied in the recurrent step, and the default function setting is hard_sigmoid, which is a piecewise linear approximation in the logistic activation function, which is used to improve the learning speed of the model to wake up faster acute hypotension prediction warning; Considering that in the training of the model, if the parameters of the model are too many and the training samples are too few, the model trained is easy to produce the phenomenon of overfitting, which leads to a larger loss function on the test data and a lower prediction accuracy, and in order to improve the accuracy, a lot of time is needed to repeatedly train the samples, so here, in order to reduce the problems of overfitting and time consumption, Dropout is used, which specifically includes stopping the work of the activation value of a certain neuron with a certain probability p in the forward propagation process of the model, thereby preventing it from relying too much on local features, and thereby making the model more general; Timestep is the input data associated with how many time series of each data before, the number of hidden layers depends on the reference value between 2N and log2N, N is the number of input layer nodes, when the number of hidden layer nodes is set to be less than log2N, the fitting effect of the network will decrease; When set to be greater than 2N, the training time will be prolonged, and it is easy to fall into a local minimum point; Batch_size represents the number of samples trained at a time, which will affect the optimization degree and training speed of the model, the larger the Batch_size value, the faster the training speed, that is, the faster the result error converges, but the generalization ability of the model will decrease, so the value is determined according to the amount of data during detection; Epochs is the number of iterations, that is, the number of complete training using all samples, set LossFunction (loss function) to MAE (mean absolute error) to represent the training result error, when the Loss function converges, stop training; Optimizer is the optimizer, which is set to Adam, which is used to calculate the update step, compared with MomentumAdagrad, Adadelta, RMSprop, etc. Optimizer, it improves the learning efficiency of the model and strengthens the adaptability. In addition, the root mean square error (RMSE) of the predicted data and the real data is calculated to measure the deviation between the predicted value and the actual value, RMSE is the root mean square error of the predicted data and the real data, which is used to evaluate the degree of change of the data, the smaller the RMSE, the higher the prediction accuracy of the prediction model, so as to adjust the parameters in the LSTM network.

Claims

1. An acute hypotension hybrid early warning method based on an LSTM network, characterized in that, The acute hypotension mixed early warning method comprises the following steps: Step A, import the collected physiological data sequence, detect the hypotension event, if the current is in the hypotension state, output the result as a hypotension early warning, enter step C, otherwise, enter step B; Step A further comprises detecting whether it has been in the acute hypotension state, the detection process comprising the following steps: Step 1, according to the imported collected physiological data sequence, fix a pointer as the starting point and the tail as the end point in the queue of the physiological data sequence, and use an integer variable N to record the number of data in the mean arterial pressure sequence whose blood pressure value is lower than the hypotension threshold from the starting point to the end point; Step 2, every time data is enqueued, check whether the mean arterial pressure of the data pointed to by the fixed pointer is lower than the hypotension threshold, if yes, the integer variable N is decremented by one, otherwise, do not add or subtract, enqueue new data; Step 3, if the new blood pressure value is less than the hypotension threshold, then the integer variable N is incremented by one, otherwise, do not add or subtract; Step 4, judge whether the ratio of the integer variable N and the reaction time is greater than the hypotension threshold percentage, if yes, output the result as a hypotension state; Step B, predict the acute hypotension event of the current collected physiological data sequence, if the next period will occur acute hypotension, output the result as a hypotension prediction early warning, and then enter step C; Step C, detect the blood pressure protection value, if it exceeds the protection value, output the result blood pressure exceeds the protection value, return to the two results in steps A and B for early warning, the blood pressure protection value is the hypotension threshold; Wherein, the prediction process in step B comprises establishing an early warning model based on an LSTM network, and training the model, the training steps comprising step X1, step X2, step X3, step X4 in turn; Step X1, import the collected physiological data sequence, and preprocess the data; Step X2, wavelet decomposition is performed on the preprocessed signal; Step X3, predict the next time detail coefficient and proximity coefficient through the LSTM neural network structure; Step X4, wavelet reconstruction is performed on the data predicted in the previous step, and then the signal time sequence prediction result is obtained; The prediction of acute hypotension event in step B is to put the imported collected physiological data sequence into the model trained in step X for prediction, which comprises the following process: For a period of time, the physiological signal data of a patient, whether the mean arterial pressure meets The mean arterial pressure calculation formula is: ; Wherein, MAP is the mean arterial pressure; Diastolic is the diastolic pressure; Systolic is the systolic pressure; p represents the hypotension threshold, whose unit is mmHg; t represents the reaction time, whose unit is minute; Q represents the hypotension threshold percentage; q represents the hypotension threshold percentage.

2. The LSTM network-based acute hypotension hybrid warning method according to claim 1, characterized in that, The data preprocessing in step X1 comprises the following steps: Step 1, if there is an outlier in the imported collected physiological data sequence, the outlier is removed, and the data with mean arterial pressure higher than 140 mmHg and lower than 35 mmHg is set to zero; Step 2, fill in the missing values, and fill in the missing values using linear interpolation, the specific formula is as follows: ； Wherein, (x1, y1) and (x0, y0) represent the end points at both ends of the missing segment. 3.The LSTM network-based acute hypotension hybrid warning method of claim 1, wherein, In step X3, the LSTM network includes an input layer, a hidden layer, and an output layer. The neurons in the hidden layer are added with output gates, input gates, and forget gates. At the same time, the parameters in the LSTM network structure are adjusted, including: Activation layer, with the default activation function set to tanh; Recurrent_activation loop step, with the default activation function set to hard_sigmoid; Add Dropout, neurons in training are discarded according to probability P, P is 0.5; Timestep setting, set to associate each data with its previous time series input data; The number of hidden layers is set to log2N~2N, N is the number of input layer nodes; Batch_size setting, set to the total number of samples in this training model; Epochs setting, which corresponds to the number of complete training of all samples; Loss Function setting, set Loss Function to MAE, when Loss Function converges, stop training the model; Optimizer setting, use Adam.

Images