Blood pressure monitoring method based on long short-term memory network model

Abstract

The invention discloses a blood pressure monitoring method based on a long short-term memory network model. According to the method, the problem of how to provide a blood pressure monitoring method which can be deeply fused with a safety helmet, adapts to the working environment and is high in accuracy is solved, and efficient and accurate blood pressure monitoring is achieved by combining data collection at a specific part with deep learning. The method specifically comprises the following steps: acquiring pulse wave signals of a testee, wherein the acquisition parts comprise intersection positions of arteries on two side surfaces of the head and mandible, temporal regions on two sides of the head and the middle part of a frontal region; filtering and feature extraction are carried out on original signals, data are integrated to construct a reference data set, the data set is trained by using a long-short-term memory network model to obtain an optimal model, and finally blood pressure analysis is carried out on the collected data by using the optimal model. By optimizing the acquisition strategy and using the long-short-term memory network model, the physiological information of the testee is acquired and monitored on the head, and the system is suitable for scenes such as intelligent safety helmets and the like.

Description

Technical Field

[0001] The embodiments of the present invention relate to the technical field of intelligent safety helmets, and relate to a method for monitoring physiological parameters of specific parts by combining data collection with deep learning, specifically a blood pressure monitoring method based on a long short-term memory network model. Background Art

[0002] As a safety device that integrates safety protection and auxiliary operations, in order to better protect the life safety of the wearer, it is necessary to realize real-time monitoring of the health status of the wearer. Blood pressure, as one of the important physiological parameters of the human body, is an important indicator for evaluating health status. Therefore, accurate monitoring of blood pressure is an essential part of intelligent safety helmets. In recent years, non-invasive continuous blood pressure monitoring technology has gradually become a research hotspot. Currently, the commonly used blood pressure monitoring methods are mainly the pulse wave transit time (PWTT) method based on photoplethysmography (PPG) and electrocardiogram (ECG) signals, and the piezoelectric sensing technology based on the radial artery or fingertips.

[0003] In the field of intelligent safety helmets, there are also many difficulties in realizing blood pressure monitoring. First of all, the existing non-invasive blood pressure monitoring methods mainly rely on sensors on the wrist and finger parts, and the detection methods on the head have not been developed enough. Secondly, compared with parts such as the wrist and finger, the signals that can be collected on the head are often weaker, and appropriate signal collection and signal processing methods need to be selected. Moreover, there is a scarcity of data sets. Most of the existing data sets are established on the signal collection of traditional parts, and there is a lack of suitable public data sets for training the model.

[0004] In summary, the technical bottlenecks of the existing blood pressure monitoring technology are as follows: First, the compatibility between traditional monitoring devices and safety helmets is poor, which affects operation safety and user experience; second, the traditional algorithms have insufficient ability to represent the signal characteristics of the head; third, the existing data sets lack signal data collected based on the head. Therefore, there is an urgent need to develop a blood pressure monitoring method that is deeply integrated with the safety helmet, adapts to the working environment, and has high accuracy. Summary of the Invention

[0005] The technical problem to be solved by the present invention is how to provide a blood pressure monitoring method that can be deeply integrated with the safety helmet, adapts to the working environment, and has high accuracy.

[0006] To solve the above technical problems, the technical solution adopted by the present invention is: A blood pressure monitoring method based on a long short-term memory network model, which is characterized by including the following steps:

[0007] A blood pressure monitoring method based on a long short-term memory network model, which is characterized by including the following steps:

[0008] Step (1): Use a photoelectric pulse wave sensor to collect the pulse wave data of the subjects under various physical conditions.

[0009] Step (2): Preprocess the data, construct a benchmark data set by integrating the data, and divide it into a training set and a validation set.

[0010] Step (3): Construct a long short-term memory network model suitable for blood pressure monitoring, train the benchmark data set obtained in step (2), and generate a model.

[0011] In step (1), male and female subjects with a recruitment ratio of approximately 2:1 are recruited, and the pulse waves, blood pressure, heart rate, etc. of the subjects in a calm state and after strenuous exercise are collected, covering diverse pulse wave data under different genders and various blood pressure conditions. The signal monitoring sites of the pulse wave are: the intersection position of the bilateral facial arteries and the mandible on both sides of the head, the bilateral temporal regions of the head, and the middle frontal region, totaling five positions.

[0012] In step (2), the data preprocessing part includes: a filtering module and a feature extraction module.

[0013] Filtering module: Used to remove the noise of the original data, smooth the data, and improve the signal quality.

[0014] A further technical solution lies in that the process of the filtering module filtering the original data includes:

[0015] Use a Butterworth filter for band-pass filtering to achieve the functions of denoising and smoothing the original data.

[0016] Feature extraction module: Used to perform highlighting processing on the filtered signal to obtain the periodic features in the signal.

[0017] A further technical solution lies in that the process of the feature extraction module extracting features from the filtered data includes:

[0018] Extract the pulse wave periods of all five channels of signals, and extract the period start point, systolic peak, dicrotic notch, and diastolic peak of each period from them.

[0019] In step (2), the part of constructing the benchmark data set includes:

[0020] In addition to the filtered data and features, it also includes the signal amplitude difference from the facial artery collection point to the temporal region collection point, the signal amplitude difference from the temporal region collection point to the middle frontal region collection point, and the physiological information of the subjects, including gender, age, height, weight, systolic blood pressure during measurement, diastolic blood pressure during measurement, and heart rate during measurement, to jointly construct the benchmark data set and divide it into a training set and a validation set.

[0021] In step (3), the process of training the long short-term memory model includes the following steps:

[0022] Step (3-1) passes the data through the LSTM layer and the fully connected layer to generate prediction results, and uses a loss function to calculate the error between the predicted value and the true value, thereby determining the impact of each parameter on the loss. After the calculation is completed, an optimizer is used to update the weights and biases of the model to make the model better fit the data.

[0023] After all the data processing in step (3-2) is completed, the performance of the model is evaluated using the validation set, and the model loss value on the validation set is used as the evaluation criterion.

[0024] Steps (3-1) to (3-2) are repeatedly executed to optimize and adjust each parameter. When the loss value no longer decreases, the optimal model is obtained, and this optimal model serves as the benchmark for the monitoring method.

[0025] The beneficial effects of the present invention are as follows:

[0026] In the blood pressure monitoring method based on the long short-term memory network model of the present invention, first, a photoelectric plethysmogram sensor is used to collect the plethysmogram data of the subject under various physical conditions. The collected plethysmogram data is filtered and feature-extracted, and combined with various information to construct a benchmark data set for blood pressure monitoring, and the training set and the validation set are divided; a long short-term memory network model LSTM suitable for blood pressure monitoring is constructed; in the established long short-term memory network model, the plethysmogram data and related information of the subject under various physical conditions collected are trained to obtain the trained optimal model; the optimal model is verified using the pre-divided validation set to detect the blood pressure monitoring ability of the model in the actual use scenario. The present invention optimizes the plethysmogram acquisition site, uses filtering and feature extraction methods suitable for blood pressure monitoring, and solves the problem that traditional monitoring methods cannot be used on the head. A new plethysmogram data set is established, providing benchmark data for the development of blood pressure monitoring methods in the head area, and being more able to meet the application requirements in the field of intelligent safety helmets compared with the existing data sets. Therefore, the present invention can solve the limitations of the existing blood pressure monitoring methods and provide an efficient and accurate method for implementing blood pressure monitoring on the head. Description of the Drawings

[0027] In order to more clearly illustrate the implementation of the present invention or the existing technical solutions, the following will briefly introduce the drawings required for the description of the embodiments or the existing technology.

[0028] Figure 1 It is a diagram of the blood pressure monitoring method

[0029] Figure 2 It is a schematic diagram of the plethysmogram acquisition position Detailed Embodiments

[0030] The main steps of the method provided by the present invention are as follows:

[0031] Step (1): Use a photoelectric pulse wave sensor to collect the pulse wave data of the subject under various physical conditions.

[0032] Step (2): Preprocess the data, construct a benchmark data set by integrating the data, and divide it into a training set and a validation set.

[0033] Step (3): Construct a long short-term memory network model suitable for blood pressure monitoring, train the benchmark data set obtained in step (2), and generate a model.

[0034] Step (4): Use the generated model to analyze new data and output the results.

[0035] As a preferred embodiment of the present invention, in step (1), the process of collecting data is as follows:

[0036] Step (1-1): Recruit male and female subjects with a ratio of about 2:1, and use a photoelectric pulse wave sensor to collect the pulse wave, blood pressure, heart rate, etc. data of the subjects under calm state and after strenuous exercise, covering diverse pulse wave data of different genders and various blood pressure conditions.

[0037] The signal acquisition sites of the pulse wave are respectively: the intersection position of the bilateral facial arteries on the head and the mandible, the bilateral temporal regions on the head, and the middle frontal region, a total of five positions.

[0038] As a preferred embodiment of the present invention, in step (2), the part of data preprocessing includes the following steps:

[0039] Step (2-1): Perform band-pass filtering using a Butterworth filter to achieve the functions of denoising and smoothing the original data.

[0040] Step (2-2): Extract the pulse wave periods of all five channels of signals, and extract the period starting point, systolic peak, dicrotic notch, and diastolic peak of each period.

[0041] As a preferred embodiment of the present invention, in step (2), the part of constructing the benchmark data set includes the following steps:

[0042] Step (2-3): Calculate the signal amplitude difference between the facial artery acquisition point and the temporal region acquisition point, and the signal amplitude difference between the temporal region acquisition point and the middle frontal region acquisition point.

[0043] Step (2-4): Collect the physiological information of the subject, including gender, age, height, weight, systolic blood pressure during measurement, diastolic blood pressure during measurement, and heart rate during measurement.

[0044] Step (2-5): Integrate the above data, construct a benchmark data set, and divide it into a training set and a validation set.

[0045] As a preferred embodiment of the present invention, in step (3), the process of training the long short-term memory model includes the following steps:

[0046] Step (3-1) Transmit data through the LSTM layer and the fully connected layer to generate a prediction result, and use a loss function to calculate the error between the predicted value and the true value, so as to determine the impact of each parameter on the loss. After the calculation is completed, use an optimizer to update the weights and biases of the model to make the model better fit the data.

[0047] Step (3-2) After all the data processing is completed, use the validation set to evaluate the performance of the model, and use the model loss value on the validation set as the evaluation criterion.

[0048] Step (3-3) Repeat steps (3-1) to (3-2) to optimize and adjust each parameter. When the loss value no longer decreases, obtain the optimal model, and use this optimal model as the benchmark for the monitoring method.

[0049] As a preferred embodiment of the present invention, in step (4), analyzing the data includes the following steps:

[0050] Step (4-1) Collect new data and make it into a data set in the manner of step (2).

[0051] Step (4-2) Use the optimal model obtained in step (3) to analyze the data and output the result.

[0052] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the following combines the appended Figure 1 of the present invention to describe the technical solutions in the embodiments of the present invention completely and clearly:

[0053] As Figure 1 shown, first collect the pulse wave signal of the subject, and the collection site is as Figure 2 shown. Then filter and extract features from the original signal, construct a benchmark data set by integrating the data, train the data set using the long short-term memory network model to obtain the optimal model, and finally use the optimal model to analyze the blood pressure of the collected data.

[0054] In the collection process of the present invention, the recruited subjects consist of 11 female subjects and 24 male subjects. Each person undergoes two groups of collections. The first group collects data five times under a calm state, each collection lasting for 1 minute. After the collection is completed, perform rapid exercise for about three minutes, and then immediately perform the second group of collections, also collecting five times, each time lasting for 1 minute.

[0055] In the filtering process of the present invention, the passband cut-off frequency of the Butterworth filter used is 0.4 Hz, and the stopband cut-off frequency is 4 Hz; in the feature extraction process, first, each pulse wave period is determined according to the extreme values to obtain the starting point of the period, the systolic peak value is obtained according to the maximum value within the period, then the second derivative of the pulse wave signal is calculated, and the dicrotic notch and diastolic peak value are obtained according to the zero points of its second derivative.

[0056] In the construction process of the reference data set of the present invention, first, the signal amplitude difference between the facial artery acquisition point and the temporal region acquisition point, and the signal amplitude difference between the temporal region acquisition point and the middle frontal region acquisition point are calculated, and then the physiological information of the subject is integrated to construct a data set of size (1,936,400, 36), where 1,936,400 is the number of pulse wave signals collected, and 36 is a total of 36 items of data including five-way signals and their features, amplitude differences, gender, age, height, weight, systolic blood pressure during measurement, diastolic blood pressure during measurement, and heart rate during measurement.

[0057] In the training process of the present invention, the network model used is a long short-term memory network model, the hidden layer size is 100, the learning rate is 0.000001, the size of one training epoch is 100, and the batch size is 14. In each training process, the model first passes the data through the LSTM layer and the fully connected layer to generate prediction results, and uses the loss function to calculate the error between the predicted value and the true value, so as to determine the influence of each parameter on the loss. After the calculation is completed, the optimizer is used to update the weights and biases of the model to make the model better fit the data. This process is continuously repeated until the loss value no longer decreases. In this embodiment, the training loss and test loss of diastolic blood pressure converge to 1.12×10 -9 and 3.43×10 -9 , the training loss and test loss of systolic blood pressure converge to 1.57×10 -4 and 2.21×10 -4 .

[0058] In the verification process of the present invention, the model is verified using the verification set that has not participated in the training. First, the obtained optimal model is used to predict the data, and then it is compared with the true value to calculate the mean square error MSE, mean absolute error MAE, root mean square error RMSE, and mean absolute percentage error MAPE with the true value. In this embodiment, the MSE of the optimal model predicting diastolic blood pressure is about 20.83, the MAE is about 3.73, the RMSE is about 4.54, and the MAPE is about 5.24%; the MSE of the optimal model predicting systolic blood pressure is about 34.02, the MAE is about 4.69, the RMSE is about 5.83, and the MAPE is about 4.34%, all meeting the American Association for the Advancement of Medical Instrumentation AAMI standard (average error <= ±5 mmHg, root mean square error <= ±8 mmHg).

Claims

1. A blood pressure monitoring method based on a long short-term memory network model, characterized in that: The following steps are involved: Step (1) using a photoelectric pulse wave sensor to collect pulse wave data of a subject under various physical conditions; Step (2) preprocessing the data, integrating the data to construct a benchmark data set, and dividing it into a training set and a validation set; Step (3) constructing a long short-term memory network model suitable for blood pressure monitoring, training the benchmark data set obtained in step (2) to generate a model; Step (4) uses the generated model to analyze the new data and output the results.

2. A blood pressure monitoring method based on a long short-term memory network model according to claim 1, characterized in that: In step (1), pulse wave, blood pressure, heart rate and other data of the volunteers are collected when they are in a calm state and after strenuous exercise, covering diverse pulse wave data of different genders and under various blood pressure conditions.

3. A blood pressure monitoring method based on a long short-term memory network model according to claim 1, characterized in that: In step (1), the pulse wave signal monitoring positions are: the intersection of the lateral arteries on both sides of the head and the mandibular bone, the temporal areas on both sides of the head, and the middle of the frontal area, a total of five positions.

4. The blood pressure monitoring method based on the long short-term memory network model according to claim 1, characterized in that: In step (2), the data preprocessing part includes: a filtering module and a feature extraction module; Among them, the filtering module is used to remove noise from the original data, smooth the data, and improve the signal quality; the feature extraction module is used to highlight the filtered signal and obtain the periodic features in the signal.

5. A blood pressure monitoring method based on a long short-term memory network model according to claim 4, characterized in that: The process of filtering the original data by the filtering module includes: The Butterworth filter is used for bandpass filtering to achieve denoising and smoothing of the original data.

6. A blood pressure monitoring method based on a long short-term memory network model according to claim 4, characterized in that: The process of feature extraction module extracting features from filtered data includes: The pulse wave cycles of all five signals are extracted, and the cycle starting point, systolic peak, dicrotic notch, and diastolic peak of each cycle are extracted from them.

7. A blood pressure monitoring method based on a long short-term memory network model according to claim 1, characterized in that: In step (2), the part of building the benchmark dataset includes: In addition to the filtered data and features, it also includes the signal amplitude difference from the facial artery acquisition point to the temporal area acquisition point, the signal assignment difference from the temporal area acquisition point to the middle frontal area acquisition point, and the physiological information of the subjects, including gender, age, height, weight, systolic blood pressure during testing, diastolic blood pressure during testing, and heart rate during testing. The benchmark data set is jointly constructed and divided into training set and validation set.

8. The blood pressure monitoring method based on the long short-term memory network model according to claim 1, characterized in that: In step (3), the process of training the long short-term memory model includes the following steps: S1, passes data through the LSTM layer and the fully connected layer to generate prediction results, and uses the loss function to calculate the error between the predicted value and the true value to determine the impact of each parameter on the loss. After the calculation is completed, the optimizer is used to update the weights and biases of the model so that the model better fits the data; S2, after all data are processed, use the validation set to evaluate the performance of the model, and use the model loss value on the validation set as the evaluation criterion; S3, repeat S1 to S2 to optimize and adjust various parameters. When the loss value no longer decreases, the optimal model is obtained, and the optimal model is used as the benchmark of the monitoring method.

Images