LU502768B1

LU502768B1 - Intelligent blood pressure prediction method based on multi-scale residual network and ppg signal

Info

Publication number: LU502768B1
Application number: LU502768A
Authority: LU
Inventors: Chengfeng Sun
Original assignee: Univ Nanjing Finance & Economics
Priority date: 2022-06-17
Filing date: 2022-07-20
Publication date: 2023-12-22
Also published as: CN115089145A; WO2023240739A2

Abstract

The present disclosure relates to the technical field of non-invasive blood pressure prediction, and specifically discloses an intelligent blood pressure prediction method based on a multi-scale residual network and a PPG signal, which includes: collecting the PPG signal of a user; filtering the collected PPG signal of the user; segmenting the filtered PPG signal, and attaching a corresponding blood pressure label to the filtered PPG signal; dividing a training set and a test set; constructing a blood pressure prediction network model based on the multi-scale residual network; inputting the training set into the blood pressure prediction network model for training; and inputting the test set into the trained blood pressure prediction network model to verify effectiveness of the blood pressure prediction network model.

Description

BL-5546 as amended

INTELLIGENT BLOOD PRESSURE PREDICTION METHOD BASED ON LUS02768

MULTI-SCALE RESIDUAL NETWORK AND PPG SIGNAL

CROSS REFERENCE TO RELATED APPLICATION

[0001] This patent application claims the benefit and priority of Chinese Patent

Application No. 202210690723.3, entitled “Intelligent blood pressure prediction method based on multi-scale residual network and PPG signal” filed on June 17, 2022, the disclosure of which is incorporated by reference herein in its entirety as part of the present application.

TECHNICAL FIELD

[0002] The present disclosure relates to the technical field of non-invasive blood pressure prediction, in particular to an intelligent blood pressure prediction method based on a multi-scale residual network and a PPG signal.

BACKGROUND ART

[0003] In order to achieve continuous, comfortable and long-term blood pressure (BP) monitoring, many sleeveless noninvasive sensors have been proposed, which can perform the long-term, continuous and comfortable BP monitoring. These sleeveless sensors can even collect long-term data in the form of wearable devices for BP estimation. Cuffless blood pressure measurement devices are being widely used to improve blood pressure measurement. Many of these devices use photoplethysmography (PPG) sensors, the PPG sensors measure pulse arterial blood volume waveforms by means of shining light on one side of a tissue volume and receiving transmitted light on the other side or reflected light on the same side.

Compared to other plethysmographs such as ultrasound, these sensors are easy to use and require minimal hardware. Traditional PPG sensors come in the form of hand clips on fingertips, toes or earlobes, and reflective patches on a nose, forehead or elsewhere.

With dedicated sensors or cameras, reflective PPG sensors can now also be used in smartphones and smartwatches, and may also become a suitable bioelectronic system.

[0004] However, even the PPG and electrocardiogram (ECG) sensors have successfully obtained pulse wave velocity (PWV), pulse transit time (PTT) and pulse arrival time (PAT), and it has never been fully demonstrated that PTT or PAT can accurately estimate

BP. This is because, to describe the dynamics of a real cardiovascular system in humans, it is not enough to consider PWV, PTT and PAT alone. Indeed, the measured PPG waveforms are highly dependent on the variation in cardiovascular characteristics among people. For these people, BP prediction is often performed by a given algorithm designed to extract many predetermined (or hand-crafted) features at different time intervals, but wrong features may be obtained. In addition, the optical properties of the skin, tissue, fat, and bone around and above the blood vessels measured by different subjects can also affect the waveform features of the PPG, not to mention the possible wrong positioning of the measurement position. With the deepening of intelligent regression research, various signal preprocessing techniques are used to extract quality features from raw measured PPG waveforms, but the raw measured PPG waveforms 1

BL-5546 as amended usually have large direct current (DC) drift and high-frequency noise. Another group of LU502768 reported studies utilizes various regression methods described above to extract non-physiological features from the PPG waveforms to estimate BP values.

[0005] Recently, some works using deep learning methods to directly extract features from PPG data to predict BP have been reported. A deep learning model that integrates feature extraction and regression calculation can better solve the problem of large BP prediction errors caused by manual feature extraction. In addition, deep learning provides an opportunity to achieve better BP accuracy by using a single PPG sensor. On this basis, how to use the deep learning method to design a high-precision and efficient

BP prediction model is the focus of the research.

SUMMARY

[0006] Based on this, it is necessary to provide an intelligent blood pressure prediction method based on a multi-scale residual network and a PPG signal, which is helpful for a blood pressure prediction network to extract features of different scales in a process of processing the PPG signal, so that the learning ability of the network is improved, and the prediction accuracy of the blood pressure prediction network is effectively improved.

[0007] In order to achieve the above object, the present disclosure provides the following solutions:

[0008] An intelligent blood pressure prediction method based on a multi-scale residual network and a PPG signal includes following steps:

[0009] S1, collecting the PPG signal of a user;

[0010] S2, filtering the collected PPG signal of the user, and obtaining the filtered PPG signal;

[0011] S3, segmenting the filtered PPG signal according to a preset sample duration, and obtaining multiple PPG signal samples;

[0012] S4, dividing the segmented multiple PPG signal samples into a training set and a test set, and normalizing PPG signal samples in the training set and PPG signal samples in the test set;

[0013] S5, constructing an initial multi-scale residual network blood pressure prediction model;

[0014] S6, inputting the normalized PPG signal samples in the training set into the initial multi-scale residual network blood pressure prediction model for training, and obtaining the trained multi-scale residual network blood pressure prediction model; and

[0015] S7, inputting the normalized PPG signal samples in the test set into the trained multi-scale residual network blood pressure prediction model for blood pressure prediction, and obtaining a blood pressure prediction result of the user.

[0016] Compared with the prior art, the beneficial effects of the present disclosure are:

[0017] In the present disclosure, a multi-scale residual network is used, more abundant feature information is extracted from the PPG signal, and then the network is deepened by using a convolution layer of 3x1 convolution kernels, thereby the generalization of a blood pressure prediction network is improved, and the prediction accuracy of the blood pressure prediction network is effectively improved. 2

BL-5546 as amended

BRIEF DESCRIPTION OF THE DRAWINGS LUs02768

[0018] In order to more clearly illustrate the embodiments of the present disclosure or the technical solutions in the prior art, the accompanying drawings required in the embodiments will be briefly introduced below. The drawings in the following description are only some embodiments of the present disclosure. For those of ordinary skill in the art, other drawings can also be obtained from these drawings without creative labor.

[0019] FIG. 1 is a flow chart of an intelligent blood pressure prediction method based on a multi-scale residual network and a PPG signal;

[0020] FIG. 2 is a specific implementation flow chart of the intelligent blood pressure prediction method based on a multi-scale residual network and a PPG signal,

[0021] FIG. 3 is a schematic diagram of an overall structure of a multi-scale residual network blood pressure prediction model,

[0022] FIG. 4 is a regression analysis diagram of a systolic blood pressure prediction result; and

[0023] FIG. 5 is a regression analysis diagram of a diastolic blood pressure prediction result.

DETAILED DESCRIPTION OF THE EMBODIMENTS

[0024] The technical solutions in the embodiments of the present disclosure will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present disclosure, and the described embodiments are only a part of the embodiments of the present disclosure, rather than all the embodiments.

Based on the embodiments of the present disclosure, all other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the protection scope of the present disclosure.

[0025] An intelligent blood pressure prediction method based on a multi-scale residual network and a PPG signal provided in the embodiment, as shown in FIGS. 1-2, includes:

[0026] S1, collecting the PPG signal of a user.

[0027] It should be noted that the PPG signal is a photoplethysmography signal.

[0028] S2, filtering the collected PPG signal of the user to remove baseline drift and high-frequency noise, and obtaining the filtered PPG signal.

[0029] S3, segmenting the filtered PPG signal according to a preset sample duration, obtaining multiple PPG signal samples, and attaching corresponding diastolic blood pressure and systolic blood pressure labels to each of the samples.

[0030] S4, dividing the segmented multiple PPG signal samples into a training set and a test set, and normalizing PPG signal samples in the training set and PPG signal samples in the test set.

[0031] S5, constructing an initial multi-scale residual network blood pressure prediction model.

[0032] S6, inputting the normalized PPG signal samples in the training set into the initial multi-scale residual network blood pressure prediction model for training, and obtaining the trained multi-scale residual network blood pressure prediction model. 3

BL-5546 as amended

[0033] It should be noted that the trained multi-scale residual network blood pressure LU502768 prediction model 1s a feature extraction model and a blood pressure prediction model for the PPG signal.

[0034] S7, inputting the normalized PPG signal samples in the test set into the trained multi-scale residual network blood pressure prediction model for blood pressure prediction, and obtaining a blood pressure prediction result of the user.

[0035] Preferably, the following steps are further included:

[0036] collecting the PPG signal of the user by means of a PPG signal collection device; and

[0037] setting the sample duration for each analysis, for example, the sample duration is 8s, and segmenting the PPG signal by a sliding window method according to the sample duration, and dividing the segmented PPG signal samples into the training set and the test set.

[0038] Preferably, the step S4 includes:

[0039] scaling the PPG signal samples in the training set and the PPG signal samples in the test set by a min-max normalization method, and reducing an amplitude of the signal, so that the training efficiency and recognition accuracy of the initial multi-scale residual network blood pressure prediction are improved.

[0040] Preferably, as shown in FIG. 3, the step SS includes:

[0041] determining hyperparameters of each module of the initial multi-scale residual network blood pressure prediction model according to structural parameters of a

Resnet-18 network, wherein the hyperparameters comprise a convolution kernel size, a channel number and a stride of each convolutional layer, a pooling size and a stride of each pooling layer, and a specific construction process of the initial multi-scale residual network blood pressure prediction model is as follows:

[0042] step S5.1, constructing 3 convolutional layers with 1x1 convolution kernels and 1 max-pooling layer after an input layer, wherein 2 convolutional layers with a 1x1 convolution kernel are followed by a convolutional layer with 3x1 convolution kernels and a convolutional layer with 5x1 convolution kernels respectively, and the max-pooling layer is followed by a convolutional layer with a 1x1 convolution kernel for capturing multi-scale features in the PPG signal;

[0043] step S5.2, then constructing a channel splicing layer, and constructing 4 convolutional layers with the channel number of 16, the convolution kernel size of 3 x 1 and the stride of 1 in turn, and adding a skip connection between two convolutional layers at the same time; wherein, two skip connections are added in total,

[0044] step S5.3, then constructing 4 convolutional layers with the channel number of 32 and the convolution kernel size of 3 x 1 in turn, wherein the stride of a first convolutional layer is 2, which is configured to compress the features, and the stride of remaining convolutional layers is 1; and adding the skip connection between last two convolutional layers at the same time;

[0045] step S5.4, then constructing 4 convolutional layers with the channel number of 64 and the convolution kernel size of 3 x 1 in turn, wherein the stride of a first convolutional layer is 2, and the stride of remaining convolutional layers is 1; and adding the skip connection between last two convolutional layers at the same time; and 4

BL-5546 as amended

[0046] step S5.5, finally constructing a global average pooling layer, and inputting the LU502768 multi-scale features of the PPG signal output by the global average pooling layer into a fully connected layer with 2 neurons to obtain the trained multi-scale residual network blood pressure prediction model.

[0047] Preferably, the normalized PPG signal samples in the training set are input into the initial multi-scale residual network blood pressure prediction model for training, wherein a learning rate is set to 0.001; and a mean square error of a true blood pressure value and a predicted blood pressure value is minimized by using an Adam optimizer.

[0048] It should be noted that the PPG signal is from data in the Multiparameter

Intelligent Monitoring in Intensive Care (MIMIC) II online waveform database provided by PhysioNet, and the datasets include normal, hypertensive and hypotensive populations.

[0049] It should be noted that the PPG signal is obtained by collecting a fingertip PPG signal through an oximeter, and the blood pressure is calculated by means of Artery

Blood Pressure (ABP).

[0050] In the embodiment of the present disclosure, the blood pressure prediction results are shown in FIGS. 4 to 5. It can be seen that a fitting line of estimated values of systolic blood pressure (SBP) and diastolic blood pressure (DBP) is very close to a fitting line of real blood pressure values, and the predict effect is good.

[0051] In an intelligent blood pressure prediction method based on a multi-scale residual network and a PPG signal provided by the present disclosure, a multi-scale residual network is used, more abundant feature information is extracted from the PPG signal, and then the network is deepened by using a convolution layer of 3x1 convolution kernels, thereby the generalization of a blood pressure prediction network is improved, and the prediction accuracy of the blood pressure prediction network is effectively improved.

[0052] The principles and implementations of the present disclosure are described herein using specific examples. The descriptions of the above embodiments are only used to help understand the method and the core idea of the present disclosure; meanwhile, for those skilled in the art, according to the idea of the present disclosure, there will be changes in the specific embodiments and application scope. In conclusion, the contents of this specification should not be construed as a limitation of the present disclosure. 5

Claims

BL-5546 as amended CLAIMS LU502768

1. An intelligent blood pressure prediction method based on a multi-scale residual network and a PPG signal, comprising following steps: S1, collecting the PPG signal of a user; S2, filtering the collected PPG signal of the user, and obtaining the filtered PPG signal; S3, segmenting the filtered PPG signal according to a preset sample duration, and obtaining multiple PPG signal samples; S4, dividing the segmented multiple PPG signal samples into a training set and a test set, and normalizing PPG signal samples in the training set and PPG signal samples in the test set; S5, constructing an initial multi-scale residual network blood pressure prediction model; S6, inputting the normalized PPG signal samples in the training set into the initial multi-scale residual network blood pressure prediction model for training, and obtaining the trained multi-scale residual network blood pressure prediction model; and S7, inputting the normalized PPG signal samples in the test set into the trained multi-scale residual network blood pressure prediction model for blood pressure prediction, and obtaining a blood pressure prediction result of the user.

2. The intelligent blood pressure prediction method based on a multi-scale residual network and a PPG signal according to claim 1, further comprising following steps: collecting the PPG signal of the user by means of a PPG signal collection device; and setting the sample duration for each analysis, and segmenting the PPG signal by a sliding window method according to the sample duration, and dividing the segmented PPG signal samples into the training set and the test set.

3. The intelligent blood pressure prediction method based on a multi-scale residual network and a PPG signal according to claim 1, wherein the dividing the segmented multiple PPG signal samples into a training set and a test set, and normalizing PPG signal samples in the training set and PPG signal samples in the test set, comprises: scaling the PPG signal samples in the training set and the PPG signal samples in the test set by a min-max normalization method.

4. The intelligent blood pressure prediction method based on a multi-scale residual network and a PPG signal according to claim 1, wherein the constructing an initial multi-scale residual network blood pressure prediction model, comprises: determining hyperparameters of each module of the initial multi-scale residual network blood pressure prediction model according to structural parameters of a Resnet-18 network, wherein the hyperparameters comprise a convolution kernel size, a channel number and a stride of each convolutional layer, a pooling size and a stride of each pooling layer, and a specific construction process of the initial multi-scale residual 6

BL-5546 as amended network blood pressure prediction model is as follows: LU502768 step S5.1, constructing 3 convolutional layers with 1x1 convolution kernels and 1 max-pooling layer after an input layer, wherein 2 convolutional layers with a 1x1 convolution kernel are followed by a convolutional layer with 3x1 convolution kernels and a convolutional layer with 5x1 convolution kernels respectively, and the max-pooling layer is followed by a convolutional layer with a 1x1 convolution kernel for capturing multi-scale features in the PPG signal; step S5.2, then constructing a channel splicing layer, and constructing 4 convolutional layers with the channel number of 16, the convolution kernel size of 3 x 1 and the stride of 1 in turn, and adding a skip connection between two convolutional layers at the same time; wherein, two skip connections are added in total; step S5.3, then constructing 4 convolutional layers with the channel number of 32 and the convolution kernel size of 3 x 1 in turn, wherein the stride of a first convolutional layer is 2, which is configured to compress the features, and the stride of remaining convolutional layers is 1; and adding the skip connection between last two convolutional layers at the same time; step S5.4, then constructing 4 convolutional layers with the channel number of 64 and the convolution kernel size of 3 x 1 in turn, wherein the stride of a first convolutional layer is 2, and the stride of remaining convolutional layers is 1; and adding the skip connection between last two convolutional layers at the same time; and step S5.5, finally constructing a global average pooling layer, and inputting the multi-scale features of the PPG signal output by the global average pooling layer into a fully connected layer with 2 neurons to obtain the trained multi-scale residual network blood pressure prediction model.

5. The intelligent blood pressure prediction method based on a multi-scale residual network and a PPG signal according to claim 1, further comprising: inputting the normalized PPG signal samples in the training set into the initial multi-scale residual network blood pressure prediction model for training, wherein a learning rate is set to

0.001; and minimizing a mean square error of a true blood pressure value and a predicted blood pressure value by using an Adam optimizer. 7