Deep learning-based iPPG blood pressure detection method

Abstract

The invention provides an iPPG blood pressure detection method based on deep learning. According to the method, a transfer learning strategy is adopted, and the problem that deep learning training cannot be completed due to scarcity of iPPG data is solved. According to the method, the model is pre-trained by using existing PPG data with abundant types and sufficient data volume, so that the model preliminarily understands the implicit relationship between the blood volume pulse wave and the blood pressure, and then the pre-trained model is finely adjusted by using self-collected iPPG data, so that the model is more adaptive to an iPPG blood pressure detection task. Before the data are input into the model, the data are classified according to a volume pulse wave waveform characteristic quantity K value and divided blood pressure intervals, and each interval is independently trained to adapt to different types of blood pressure detection. During actual deployment and use, a personalized model is automatically allocated to carry out subsequent blood pressure detection on a subject by taking a true value measured for the first time of the subject as a benchmark.

Description

Technical Field

[0001] The present invention relates to the technical field of non-contact blood pressure detection, and specifically to an iPPG blood pressure detection method based on deep learning. Background Art

[0002] Cardiovascular disease (CVD) is the disease with the largest number of patients in the world and has become the main reason for shortening human life span. Hypertension is one of the most common cardiovascular diseases and the main risk factor for the onset of cardiovascular diseases. In the early stage of hypertension, there are often no symptoms. Blood pressure is one of the important health indicators of the human body. Increasing the frequency of blood pressure monitoring is a good preventive measure. Traditional mercury sphygmomanometers and cuff-type electronic sphygmomanometers commonly used in homes are the most commonly used blood pressure detection devices. Although mercury sphygmomanometers have high accuracy, they are complicated to operate, require high professionalism, and mercury is toxic, so they are gradually being eliminated. Electronic sphygmomanometers have replaced them as the first choice for home monitoring, but electronic sphygmomanometers have many limitations, such as skin discomfort caused by cuffs and the inability to perform continuous measurements. For patients with skin injuries and infectious diseases, cuffs also increase the risk of contact infection. Therefore, people hope to find a more convenient, non-contact, and sustainable detection method to monitor blood pressure.

[0003] Photoplethysmography (PPG) and imaging photoplethysmography (iPPG) are both biomedical monitoring technologies based on optical measurement principles, mainly used to evaluate blood flow and pulse wave changes. They differ in hardware implementation, signal acquisition methods, and application scenarios. PPG devices consist of a light source and a photodetector, and are usually measured by contacting the skin. iPPG devices use imaging devices to take pictures without contacting the skin, relying on cameras to capture color changes on the skin surface.

[0004] Early cuffless blood pressure estimation methods used pressure pulse wave transmission time (PTT) or further calculated pulse wave velocity (PWV) to derive blood pressure through parameterized models, which had extremely high requirements for data acquisition and parameter derivation. In order to simplify multi-sensor acquisition and highly complex parameter reasoning, some people proposed to use machine learning or neural networks to estimate blood pressure based on the morphological features of PPG waveforms. Since the waveform quality is greatly affected by the sensor and noise interference, it is very difficult to manually extract a large number of waveform features on low-quality waveforms, so an end-to-end blood pressure estimation method was derived. The end-to-end blood pressure estimation method uses the time-frequency characteristics of the PPG signal and its waveform as the input of the neural network, and the feature extraction work is implemented by the model. After training, the blood pressure is directly obtained. PPG blood pressure estimation has achieved considerable progress compared with previous technologies, but it still cannot get rid of the limitations of contact. Based on the development of PPG, the birth of iPPG technology provides a possibility for non-contact blood pressure detection.

[0005] At present, there are few blood pressure studies based on iPPG technology, lack of data sets, and room for improvement in measurement accuracy. Most studies extract pulse waves based on the facial areas of a single or multiple subjects, and do not consider the impact of differences in the health of the subjects' cardiovascular system on blood pressure detection from a macro level. Therefore, the present invention proposes an iPPG blood pressure detection method based on deep learning, which is based on the iPPG signal of the wrist and fully considers the differences in the health of the subjects' cardiovascular system to improve the accuracy of non-contact blood pressure detection. Summary of the invention

[0006] The purpose of the present invention is to provide an iPPG blood pressure detection method based on deep learning in order to solve the above-mentioned problems.

[0007] The technical solution adopted by the present invention is as follows: an iPPG blood pressure detection method based on deep learning, the detection method comprising the following steps:

[0008] S1: Classify the PPG database with rich types and sufficient quantity according to the volume pulse wave waveform feature K value and blood pressure range;

[0009] S2: Apply cmor wavelet to the classified PPG waveform sequence to perform one-dimensional continuous wavelet transform to obtain a time-frequency diagram;

[0010] S3: Input the time-frequency graph into the convolutional neural network for feature extraction and blood pressure mapping to obtain multiple sets of PPG pre-trained blood pressure models;

[0011] S4: classifying the self-collected wrist iPPG signal according to the volume pulse waveform feature K value and blood pressure range;

[0012] S5: Apply cmor wavelet to perform one-dimensional continuous wavelet transform on the classified iPPG waveform sequence to obtain a time-frequency diagram;

[0013] S6: The iPPG time-frequency graph is used as the input of the corresponding pre-trained model, the feature extraction layer of the pre-trained model is frozen, and only the fully connected layer is fine-tuned to obtain multiple sets of blood pressure models that have been fine-tuned;

[0014] S7: When the subject takes the first measurement, he / she provides the true blood pressure value as a benchmark for calibrating the personalized model;

[0015] S8: Input the subject's iPPG signal and the model estimates the blood pressure value.

[0016] In a preferred embodiment, in step S1, the volume pulse wave waveform characteristic K value is a characteristic based on the change in the pulse area, which can describe the average characteristics of the pulse wave from a macroscopic level and represent the changes in some of the most important physiological parameters in the human cardiovascular system, such as peripheral vascular resistance and vascular wall sclerosis. The calculation formula for the volume pulse wave waveform characteristic K value is as follows:

[0017]

[0018]

[0019] Among them, Q m is the average pulse blood flow of the volume pulse wave; Q(t) is the volume pulse blood flow curve; Q max is the maximum value of the blood flow curve; Q min is the minimum value of the blood flow curve; T is the volume pulse wave cardiac cycle.

[0020] In a preferred embodiment, in the steps S1 and S4, since the K value mainly reflects the area of ​​the pulse graph during the cardiac cycle, a complete pulse wave is usually composed of three peaks, namely the main wave, the tidal wave and the dicrotic wave, and the cardiovascular details reflected by the peak morphology cannot be reflected from the perspective of area alone. In order to compensate for this, another classification is performed according to blood pressure, and a total of 18 subsets are divided.

[0021] In a preferred embodiment, in steps S2 and S5, the time-frequency diagram obtained by one-dimensional continuous wavelet transform can clearly describe the relationship between the change of signal frequency and time, and can also express the corresponding amplitude. The information representation of the time-frequency diagram in the time domain and frequency domain can help the model better capture the characteristics of non-stationary signals.

[0022] In a preferred embodiment, in step S3, the convolutional neural network consists of 5 two-dimensional convolutional layers and 1 fully connected layer. The convolution kernel sizes corresponding to the 5 convolutional layers are 7, 3, 3, 3, 3, respectively, wherein pooling layers are set in the 1st, 2nd, and 5th layers, and each layer is configured with a batch normalization layer. After pre-training of PPG data, the model preliminarily establishes the corresponding relationship between blood volume pulse wave and blood pressure.

[0023] In a preferred embodiment, in step S6, the fully connected layer is fine-tuned using iPPG data to achieve transfer learning from PPG estimated blood pressure to iPPG estimated blood pressure.

[0024] In a preferred embodiment, in step S7, the system uses the iPPG signal and the true blood pressure value of the subject measured for the first time as a reference, and matches the most suitable model number from the multiple sets of trained blood pressure models for subsequent blood pressure measurements of the subject.

[0025] In summary, due to the adoption of the above technical solution, the beneficial effects of the present invention are:

[0026] 1. Adopt transfer learning strategy to solve the problem that iPPG data is scarce and cannot complete deep learning training. Make full use of existing PPG data with rich types and sufficient data volume to pre-train the model, so as to achieve the initial establishment of the relationship between blood volume pulse wave and blood pressure, and then use the self-collected wrist iPPG data to fine-tune the pre-trained model to achieve transfer learning, so that the model can better adapt to the task of iPPG blood pressure detection.

[0027] 2. During training, the data is classified according to the volume pulse wave waveform feature K value and the divided blood pressure intervals, and each interval is trained independently to adapt to different types of blood pressure detection. During actual deployment and use, the iPPG signal and the true blood pressure value of the subject measured for the first time are used as a benchmark to automatically assign a personalized model for subsequent blood pressure detection. The method proposed in the present invention improves the accuracy of non-contact blood pressure detection. BRIEF DESCRIPTION OF THE DRAWINGS

[0028] Figure 1 A block diagram of the scheme design of the present invention. DETAILED DESCRIPTION

[0029] In order to make the purpose, technical solution and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not used to limit the present invention.

[0030] Reference Figure 1 , an iPPG blood pressure detection method based on deep learning, the process is as follows:

[0031] Step 1. Classify the PPG data with rich types and sufficient data according to the volume pulse wave waveform feature K value and blood pressure range. PPG is a photoplethysmography method that sends and receives light signals through a finger clip device to monitor the periodic pulsation of blood at the fingertips, and the measured waveform refers to the blood volume pulse wave. The PPG data used comes from the Multi-Parameter Intelligent Monitoring of Intensive Care (MIMIC) III provided by PhysioNet. The volume pulse wave waveform feature K value is a feature based on the change in the area of ​​the pulse graph. It can describe the average characteristics of the pulse wave from a macroscopic level and represent some of the most important physiological parameters in the human cardiovascular system, such as changes in peripheral vascular resistance and the degree of vascular wall hardening. The calculation formula for the volume pulse wave waveform feature K value is as follows:

[0032]

[0033]

[0034] Among them, Q m is the average pulse blood flow of the volume pulse wave; Q(t) is the volume pulse blood flow curve; Q max is the maximum value of the blood flow curve; Q min is the minimum value of the blood flow curve; T is the cardiac cycle of the volume pulse wave. The K value mainly reflects the situation of the pulse diagram area within the cardiac cycle. A complete pulse wave usually consists of three peaks: the main wave, the tidal wave, and the dicrotic wave. Cardiovascular details reflected by the peak morphology cannot be reflected solely from the perspective of area. To make up for this, a classification is carried out again based on blood pressure. Among them, the K value is divided into three intervals: K < 0.4, 0.4 < K < 0.5, and K > 0.5; blood pressure is divided into six intervals: isolated diastolic hypotension, hypotension, normal blood pressure, high-normal blood pressure, hypertension, and isolated systolic hypertension according to the blood pressure standards of the National Health Commission and data characteristics. A total of 18 subsets are divided.

[0035] Step 2: Perform wavelet transform on the classified PPG signal, and use the cmor wavelet with a bandwidth frequency of 1.5 and a central frequency of 1.0 as the benchmark function for one-dimensional continuous wavelet transform to obtain the time-frequency diagram of the PPG waveform sequence.

[0036] Step 3: Utilize the feature extraction ability of the convolutional neural network to obtain the abstract feature quantities of the PPG signal time-frequency diagram, and initially establish the corresponding relationship between the blood volume pulse wave and blood pressure. After each subset is trained separately, 18 blood pressure sub-models pre-trained with PPG are obtained. The entire training process uses Adam as the optimizer, uses the mean squared error loss function, and sets the learning rate to 1e-6.

[0037] Step 4: Classify the self-collected iPPG signal in the same way as the PPG signal. The collection site of the self-collected iPPG signal is the wrist. Since there are differences in quantity and type between the iPPG signal and the PPG signal itself, it is impossible to fine-tune each sub-interval. Therefore, the subsequent fine-tuning operation only targets the existing iPPG intervals.

[0038] Step 5: Perform wavelet transform on the classified iPPG signal, and use the cmor wavelet with a bandwidth frequency of 1.5 and a central frequency of 1.0 as the benchmark function for one-dimensional continuous wavelet transform to obtain the time-frequency diagram of the iPPG waveform sequence.

[0039] Step 6: Input each subset into the corresponding model. All neural network layers except the last fully connected layer of each sub-model are frozen. The entire fine-tuning process uses Adam as the optimizer, uses the mean squared error loss function, and sets the learning rate to 1e-1. Among them, for the intervals without iPPG data, the corresponding PPG pre-trained model is used as a backup. After partial fine-tuning, 18 fine-tuned sub-models are obtained.

[0040] Step 7: Before officially using this method for blood pressure testing, the subject should be calibrated for the first time to determine the matching model number. Step 8: The model inputs the subject's iPPG signal to obtain the blood pressure value.

[0041] Table 1 shows the blood pressure estimation error of the PPG pre-trained model for the self-collected iPPG signal. The essence of PPG signals and iPPG signals is to represent the blood volume pulse wave, but due to differences in measurement equipment, measurement location, and noise, the signal-to-noise ratio of the iPPG signal is worse than that of the PPG signal. As can be seen from Table 1, although the PPG signal and the iPPG signal represent blood volume pulse waves of different intensities in different locations, transfer learning can also be achieved.

[0042] Table 1

[0043]

[0044] Table 2 shows the blood pressure estimation error of the self-collected iPPG signal after the PPG pre-trained model was fine-tuned by the iPPG part. Data analysis shows that although only some sub-models were fine-tuned, various error indicators have been significantly improved. This proves the reliability and superiority of the method proposed in this paper.

[0045] Table 2

[0046]

[0047] In summary, in the present invention, a transfer learning strategy is adopted to solve the problem that iPPG data is scarce and cannot complete deep learning training. The existing PPG data with rich types and sufficient data volume is fully utilized to pre-train the model, so as to achieve the preliminary establishment of the relationship between blood volume pulse wave and blood pressure by the model, and then the self-collected iPPG data is used to fine-tune the pre-trained model to achieve transfer learning, so that the model can better adapt to the task of iPPG blood pressure detection.

[0048] In the present invention, the data is classified according to the K value of the volume pulse wave waveform feature and the divided blood pressure intervals, and each interval is trained independently to adapt to different types of blood pressure detection. In actual deployment and use, the first measured true value of the subject is used as a reference, and a personalized model is automatically assigned to perform subsequent blood pressure detection. The method proposed in the present invention effectively solves the problem that iPPG data is scarce and difficult to perform deep learning training, and improves the accuracy of non-contact blood pressure detection.

[0049] It should be noted that, in this article, the terms "include", "comprises" or any other variants thereof are intended to cover non-exclusive inclusion, so that a process, method, article or device including a series of elements includes not only those elements, but also other elements not explicitly listed, or includes elements inherent to such process, method, article or device. In the absence of further restrictions, an element defined by the sentence "comprises a ..." does not exclude the presence of other identical elements in the process, method, article or device including the element.

[0050] The above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit the same. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that the technical solutions described in the aforementioned embodiments may still be modified, or some of the technical features may be replaced by equivalents. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. An iPPG blood pressure detection method based on deep learning, characterized in that: The detection method comprises the following steps: S1: Classify the PPG database with rich types and sufficient quantity according to the volume pulse wave waveform feature K value and blood pressure range; S2: Apply cmor wavelet to the classified PPG waveform sequence to perform one-dimensional continuous wavelet transform to obtain a time-frequency diagram; S3: Input the time-frequency graph into the convolutional neural network for feature extraction and blood pressure mapping to obtain multiple sets of PPG pre-trained blood pressure models; S4: classifying the self-collected wrist iPPG signal according to the volume pulse waveform feature K value and blood pressure range; S5: Apply cmor wavelet to perform one-dimensional continuous wavelet transform on the classified iPPG waveform sequence to obtain a time-frequency diagram; S6: The iPPG time-frequency graph is used as the input of the corresponding pre-trained model, the feature extraction layer of the pre-trained model is frozen, and only the fully connected layer is fine-tuned to obtain multiple sets of blood pressure models that have been fine-tuned; S7: When the subject takes the first measurement, he / she provides the true blood pressure value as a benchmark for calibrating the personalized model; S8: Input the subject's iPPG signal and the model estimates the blood pressure value.

2. The iPPG blood pressure detection method based on deep learning as claimed in claim 1, characterized in that: In step S1, the volume pulse wave waveform characteristic K value is a characteristic based on the change of the pulse graph area, which can describe the average characteristics of the pulse wave from a macroscopic level and represent the changes in some of the most important physiological parameters in the human cardiovascular system, such as peripheral vascular resistance and vascular wall sclerosis. The calculation formula of the volume pulse wave waveform characteristic K value is as follows: Among them, Q m is the average pulse blood flow of the volume pulse wave; Q(t) is the volume pulse blood flow curve; Q max is the maximum value of the blood flow curve; Q min is the minimum value of the blood flow curve; T is the volume pulse wave cardiac cycle.

3. The iPPG blood pressure detection method based on deep learning as claimed in claim 1, characterized in that: In the steps S1 and S4, since the K value mainly reflects the area of ​​the pulse graph during the cardiac cycle, a complete pulse wave is usually composed of three peaks: the main wave, the tidal wave and the dicrotic wave. The cardiovascular details reflected by the peak shape cannot be reflected from the perspective of area alone. In order to compensate for this, another classification is performed according to blood pressure, which is divided into 18 subsets in total.

4. The iPPG blood pressure detection method based on deep learning as claimed in claim 1, characterized in that: In the steps S2 and S5, the time-frequency diagram obtained by one-dimensional continuous wavelet transform can clearly describe the relationship between the signal frequency and time, and can also express the corresponding amplitude. The information representation of the time-frequency diagram in the time domain and frequency domain can help the model better capture the characteristics of non-stationary signals.

5. The iPPG blood pressure detection method based on deep learning as claimed in claim 1, characterized in that: In step S3, the convolutional neural network consists of 5 two-dimensional convolutional layers and 1 fully connected layer. The convolution kernel sizes corresponding to the 5 convolutional layers are 7, 3, 3, 3, 3, respectively, wherein pooling layers are set in the 1st, 2nd, and 5th layers, and each layer is configured with a batch normalization layer. After pre-training of PPG data, the model preliminarily establishes the corresponding relationship between blood volume pulse wave and blood pressure.

6. The iPPG blood pressure detection method based on deep learning as claimed in claim 1, characterized in that: In step S6, the fully connected layer is fine-tuned using the iPPG data to achieve transfer learning from PPG estimated blood pressure to iPPG estimated blood pressure.

7. The iPPG blood pressure detection method based on deep learning as claimed in claim 1, characterized in that: In step S7, the system uses the iPPG signal and the true blood pressure value of the subject measured for the first time as a reference, and matches the most suitable model number from the multiple sets of trained blood pressure models for subsequent blood pressure measurements of the subject.

Images