Method for reconstructing electrocardiosignal through photoelectric volume pulse wave based on deep learning

Abstract

The invention discloses a method for reconstructing electrocardiosignals through photoelectric volume pulse waves based on deep learning, and aims at solving the problems that a traditional ECG collection mode is complex, an electrode is uncomfortable to wear, and an existing signal mapping method is low in precision. PPG signals are collected through a portable device and subjected to denoising and normalization processing, a generative adversarial network is constructed, a generator maps the PPG signals into ECG signals, and a discriminator evaluates the authenticity of the generated signals. The generator loss function is combined with the mean square error and the adversarial loss, and the discriminator loss function optimizes the identification degree of the signal; a multi-scale loss function of a time domain and a frequency domain is introduced, the consistency of the generated signals in dynamic characteristics and spectral characteristics is improved, and finally, the generated ECG signals are transmitted back to the portable equipment to be displayed. Real-time signal generation can be achieved on portable equipment such as an intelligent bracelet and a watch, the generated ECG signals have high fidelity, PQRST waveforms can be accurately restored, and the method is suitable for heart rate and heart rhythm variability analysis and abnormal heart rhythm detection.

Description

Technical Field

[0001] The present invention relates to the technical field of medical and health monitoring, and more particularly, to a method for reconstructing electrocardiogram signals based on photoplethysmography using deep learning. Background Art

[0002] Photoplethysmography (PPG) and electrocardiogram (ECG) are two important physiological signals in the field of modern biomedicine. The PPG signal collects the light intensity changes of blood flow through a photoelectric sensor to reflect cardiovascular activities, while the ECG signal collects the waveform of cardiac electrophysiological activities through skin electrodes and is the gold standard for cardiovascular health assessment. In recent years, with the wide application of wearable devices, the PPG signal has become the mainstream technology for health monitoring due to its simple acquisition, low cost, and portable device. However, the role of the PPG signal in clinical diagnosis is far less rich and accurate than that of the ECG signal, and the functional gap between the two limits the further application of the PPG.

[0003] Advantages and limitations of the PPG signal: Advantages: The PPG signal can be acquired through a simple photoelectric sensor, with low device cost and convenient operation, making it suitable for use in wearable devices. Its main applications include heart rate, blood oxygen saturation monitoring, and sleep analysis. Due to the wearable nature of the acquisition device, the PPG signal is more suitable for continuous and long-term monitoring of the user's physiological state.

[0004] Limitations: The physical mechanism of the PPG signal determines that it can only reflect the dynamic information of blood flow and cannot capture the specific characteristics of cardiac electrophysiological activities, such as P waves, QRS complexes, and T waves. Especially in the diagnosis of cardiovascular diseases, the diagnostic value of the PPG signal is limited. In addition, the PPG signal is easily affected by motion artifacts, environmental light interference, skin color, and thickness changes, resulting in insufficient signal quality and stability.

[0005] Characteristics and challenges of the ECG signal: Characteristics: The ECG signal records the dynamic changes of cardiac electrical activities and is the main basis for detecting cardiovascular diseases such as arrhythmias and myocardial infarctions. Its waveform includes characteristics such as P waves, QRS complexes, and T waves, which are a direct manifestation of the cardiac function state. The high diagnostic value of the ECG signal makes it an indispensable monitoring means in hospitals and telemedicine.

[0006] Challenge: Traditional ECG acquisition requires multiple-lead electrodes to contact the skin, which not only increases the hardware complexity of the device but also poses high requirements for user operation. Especially in daily scenarios, ECG acquisition devices are large and inconvenient to operate, making them unsuitable for long-term and continuous monitoring. In addition, poor electrode contact and external interference during the acquisition process can both lead to a decline in signal quality.

[0007] Existing research on PPG-to-ECG mapping: Linear statistical methods: Some studies have attempted to derive heart rate or heart rate variability (HRV) from PPG signals through statistical methods to approximately reflect cardiac activity. These methods are based on local waveform features of PPG signals, such as peak intervals or frequency components, and can achieve limited information extraction but cannot reconstruct the complete ECG waveform.

[0008] Traditional machine learning methods: Some studies have used traditional machine learning models such as support vector machines (SVMs) and random forests for PPG-to-ECG mapping. However, these methods have limited ability to capture the complex non-linear relationship between PPG and ECG and cannot achieve high-fidelity signal reconstruction. In addition, the deficiencies in handling temporal features of these methods result in the generated ECG signals being unable to dynamically respond to actual cardiac electrical activities.

[0009] Existing deep learning methods: In recent years, deep learning has been introduced into PPG-to-ECG signal mapping, mainly using convolutional neural networks (CNNs) to extract features from PPG signals and generate approximate ECG signals. However, due to insufficient model training and unreasonable loss function design, the generated signal quality is poor, especially in complex scenarios (such as exercise and emotional fluctuations), where the reconstructed ECG signals show significant errors.

[0010] The existing technologies have the following limitations in the combined application of PPG signals and ECG signals: PPG signals can only reflect the dynamic information of blood flow, lack direct description of cardiac electrophysiological activities, and cannot provide key electrocardiogram features such as P waves and QRS complexes.

[0011] The signals are vulnerable to external interference. Especially in exercise scenarios, the signal noise increases significantly and the reliability decreases.

[0012] Traditional ECG acquisition devices require multiple-lead patches and complex hardware support, with cumbersome operations, imposing a certain burden on users.

[0013] Not suitable for portable and long-term monitoring applications, restricting their promotion in daily health management.

[0014] Technical bottlenecks in PPG-to-ECG mapping: Insufficiency of simple models: Linear models and traditional machine learning methods are difficult to capture the complex non-linear relationship between PPG and ECG. The generated signals can only retain some statistical features and lack dynamic change information.

[0015] Poor generation quality: The existing deep learning methods are still insufficient in the generation quality of signals. Especially under noise interference, the difference between the generated signals and the real signals is significant, making it difficult to meet the actual application requirements.

[0016] Lack of multi-dimensional optimization: The existing methods mainly focus on time-domain optimization, ignoring the importance of frequency-domain information, resulting in a large difference in the spectral characteristics between the generated signals and the real signals.

[0017] Therefore, we urgently need to design a method for reconstructing electrocardiogram signals from photoplethysmogram based on deep learning to solve the above problems. Summary of the Invention

[0018] The purpose of the present invention is to solve the technical problems raised in the above background technology and provide a method for reconstructing electrocardiogram signals from photoplethysmogram based on deep learning.

[0019] The above object of the present invention is achieved as follows: On the one hand, the solution of the present invention provides a method for reconstructing electrocardiogram signals from photoplethysmogram based on deep learning, and the method includes the following steps: 1) Signal acquisition: Collect photoplethysmogram signals through a wearable device, and at the same time collect a small amount of electrocardiogram signals for model training; 2) Data preprocessing: Denoise and normalize the collected PPG signals and ECG signals. The normalization formula for PPG signals is: Where: : The normalized PPG signal; : The original PPG signal; : The mean value of the original PPG signal; : The standard deviation of the original PPG signal; 3) Deep learning modeling: Design a deep learning model based on a generative adversarial network, including a generator and a discriminator. Specifically, it includes: Generator: Used to convert PPG signals into ECG signals, and its conversion formula is: Where: : The generated ECG signal; Generator network; : The normalized PPG signal; Discriminator: It is used to evaluate the authenticity of the generated ECG signals and the real ECG signals, and optimize the generator and discriminator based on the adversarial training mechanism; 4) Loss function design: The loss function of the generator includes the mean square error between the generated signal and the real signal and the adversarial loss, which is defined as: Where: : Real ECG signal; : Generated ECG signal; Weight coefficient; Discriminator network; The loss function of the discriminator is defined as: Where: Loss function of the discriminator; : Real ECG signal; : Discriminant result of the discriminator for the generated signal; 5) Model training: Based on the PPG and ECG signal pairs, optimize the parameters of the generator and discriminator through supervised learning, and train the deep learning model to make the generated ECG signal close to the real signal; 6) Model deployment and application: Deploy the trained model in a portable device, collect PPG signals in real time, and generate ECG signals through the generator for health monitoring.

[0020] Furthermore, in step 2), band-pass filtering is performed on the PPG signal and the ECG signal for denoising, and the passband range of the filter is 0.5 Hz to 40 Hz.

[0021] Furthermore, the generator in step 3) adopts a convolutional neural network combined with a long short-term memory network architecture, and its structure is: Where CNN : PPG signal processed by the convolutional neural network; LSTM( Is the long short-term memory network module; : Parameters of the convolutional neural network; : Parameters of the long short-term memory network.

[0022] Furthermore, the output layer of the generator uses a transposed convolution operation to upsample the signal, and its formula is: Where: Intermediate feature representation of the generator; Deconv Transposed convolution operation; : Parameters of the transposed convolution operation.

[0023] Furthermore, the supervised training in step 4) uses a multi-scale loss function, which is defined as: Where: : Total loss function; : Power spectral density difference loss in the frequency domain; Weight coefficient.

[0024] Furthermore, the frequency domain loss is defined as: Where: : Power spectral density of the true ECG signal; : Power spectral density of the generated ECG signal; Frequency.

[0025] Furthermore, real-time inference of the deep learning model is implemented through embedded hardware.

[0026] Furthermore, the sampling frequency of the PPG signal is 100 Hz to 200 Hz, and the sampling frequency of the generated ECG signal is 300 Hz to 500 Hz.

[0027] In summary, compared with the prior art, the present invention has the following beneficial effects: The present invention utilizes the generative adversarial network (GAN) architecture to effectively solve the problem of insufficient information in traditional PPG signals. By extracting the complex non-linear relationship between PPG and ECG through deep learning, the generated ECG signal has high fidelity and can accurately restore electrocardiogram features such as P waves, QRS complexes, and T waves. In addition, through the optimization of the multi-scale loss function (combining time domain and frequency domain information), the generated signal is close to the real signal in both dynamic characteristics and spectral characteristics, meeting the clinical diagnosis requirements.

[0028] The present invention can achieve high-quality ECG signal reconstruction only by relying on a simple PPG signal acquisition device, avoiding the much more complex multi-lead electrodes and high-cost hardware requirements of traditional ECG acquisition devices. Through the combination of portable sensors and deep learning models, users can achieve long-term monitoring of heart health and abnormal warning without complex operations, significantly improving the applicability of the technology.

[0029] By introducing denoising preprocessing and adversarial learning mechanisms, the present invention enhances the robustness of the model to noises such as motion artifacts and environmental interference, making it applicable to various complex scenarios (such as exercise, emotional fluctuations). At the same time, the lightweight deep learning model can run in real time on embedded devices such as smart bracelets and watches, supporting low-power health monitoring requirements, and providing an efficient solution for portable healthcare and remote health management. BRIEF DESCRIPTION OF THE DRAWINGS

[0030] To more clearly illustrate the technical solutions in the present invention or the prior art, the following will briefly introduce the drawings required for use in the description of the embodiments or the prior art. Obviously, the drawings in the following description are some embodiments of the present invention. For those of ordinary skill in the art, without creative efforts, other drawings can also be obtained based on these drawings.

[0031] Figure 1 It is a flowchart of the method for reconstructing electrocardiogram signals from photoplethysmogram based on deep learning of the present invention. Specific embodiments

[0032] To make the objectives, technical solutions, and advantages of the present invention clearer, the following will clearly and completely describe the technical solutions in the present invention in conjunction with the drawings in the present invention. Obviously, the described embodiments are some, but not all, of the embodiments of the present invention. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts fall within the scope of protection of the present invention.

[0033] The following combines the attached Figure 1 drawings and multiple embodiments to describe the specific embodiments of the present invention in detail.

[0034] Embodiment 1: Overall method flow; Signal acquisition: Collect photoplethysmogram (PPG) signals and a small amount of electrocardiogram (ECG) signals through wearable devices (such as smart bracelets or watches).

[0035] The PPG signal is the change in the reflected or transmitted light intensity obtained through a photoelectric sensor, and the sampling frequency is set to 100 Hz to 200 . The ECG signal is the electrocardiogram activity obtained through skin surface electrodes, and the sampling frequency is 300 Hz to 500 Hz.

[0036] Data preprocessing: Denoise the original PPG signal using a band-pass filter to set the signal passband range to 0.5 Hz to 40 Hz to filter out high-frequency and low-frequency noise.

[0037] The normalization processing formula is: where: : The normalized PPG signal; : The original PPG signal; The mean of the PPG signal; The standard deviation of the PPG signal.

[0038] Deep learning model design: Generator design: The generator adopts a combined architecture of convolutional neural network (CNN) and long short-term memory network (LSTM) to convert PPG signals into ECG signals .

[0039] The generation formula is: Where: : The generated ECG signal; Generator model ; Discriminator design: The discriminator is used to evaluate the authenticity of the generated ECG signals and the real ECG signals .

[0040] Loss function design: Loss function of the generator: Where: Loss of the generator; : Real ECG signal; : Generated ECG signal; Weight coefficient of adversarial loss.

[0041] Loss function of the discriminator: Where: : Output of the discriminator for real ECG signals; : Output of the discriminator for generated signals.

[0042] Model training: The generator and the discriminator are trained in an adversarial learning manner to optimize the loss functions and .

[0043] At the same time, multi-scale loss functions in the time domain and frequency domain are introduced: Where: : Mean squared error in the time domain; : Frequency domain power spectrum difference loss; Weight coefficient, used to balance the optimization objectives in the time domain and frequency domain.

[0044] Model deployment: The trained model is embedded in a portable device (such as a smart bracelet) to collect PPG signals in real time , and generate corresponding ECG signals Embodiment 2: Real-time health monitoring in a smart bracelet; Hardware platform: Implement deep learning inference using an embedded processor (such as an ARM Cortex-M series chip).

[0045] Data acquisition: The bracelet is equipped with a PPG sensor that collects 100 data points per second. At the same time, a small amount of ECG data is synchronized via Bluetooth for model calibration.

[0046] Model inference: Through the optimized generator model , convert the real-time PPG signal into a high-fidelity ECG signal.

[0047] Output and application: The generated ECG signal can be used for the following functions: real-time heart rate and heart rate variability (HRV) analysis; detection and warning of abnormal heart rhythms (such as atrial fibrillation).

[0048] Embodiment 3: Multi-user remote medical application; Scenario: Used for home medical devices to collect PPG signals of family members and reconstruct ECG signals, which are transmitted to a remote server via Wi-Fi.

[0049] Model training optimization: For the individual differences of different users, adopt a transfer learning strategy and use the personalized data of the user to fine-tune the parameters of the generator .

[0050] Server-side analysis: The remote server processes the reconstructed ECG signal in real time and provides a health report in combination with doctor's advice.

[0051] Core optimization: Add a frequency-domain loss function , further optimize the spectral characteristics of the signal, and make the generated ECG signal closer to the real signal.

[0052] Formula refinement: Where: : The power spectral density of the real ECG signal; : The power spectral density of the generated ECG signal; Frequency.

[0053] Result verification: Compare the spectral characteristics before and after optimization to verify the improvement in the accuracy of the generated signal in the frequency domain.

[0054] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it; although the present invention has been described in detail with reference to the foregoing embodiments, those of ordinary skill in the art should understand that they can still modify the technical solutions described in the foregoing embodiments, or perform equivalent replacements for some of the technical features; and these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for reconstructing electrocardiogram signals from photoplethysmogram based on deep learning, characterized in that, The method includes the following steps: 1) Signal acquisition: Collect photoplethysmogram (PPG) signals through a wearable device, and at the same time collect a small amount of electrocardiogram (ECG) signals for model training; 2) Data preprocessing: Denoise and normalize the collected PPG signals and ECG signals. The normalization formula for PPG signals is: Wherein: : The normalized PPG signal; : The original PPG signal; : The mean value of the original PPG signal; : The standard deviation of the original PPG signal; 3) Deep learning modeling: Design a deep learning model based on a generative adversarial network, including a generator and a discriminator. Specifically: Generator: Used to convert PPG signals into ECG signals, and its conversion formula is: Wherein: : the generated ECG signal; generator network; : the normalized PPG signal; Discriminator: Used to evaluate the authenticity of the generated ECG signals and the real ECG signals, and optimize the generator and discriminator based on the adversarial training mechanism; 4) Loss function design: The generator loss function includes the mean square error between the generated signal and the real signal and the adversarial loss, which is defined as: Wherein: : real ECG signal; : generated ECG signal; weight coefficient; discriminator network; The discriminator loss function is defined as: Wherein: The loss function of the discriminator; : The true ECG signal; : The discrimination result of the discriminator on the generated signal; 5) Model training: Based on the pairs of PPG and ECG signals, optimize the parameters of the generator and discriminator through supervised learning, and train the deep learning model to make the generated ECG signals close to the real signals; 6) Model deployment and application: Deploy the trained model in a portable device, collect PPG signals in real time, and generate ECG signals through the generator for health monitoring purposes.

2. The method for reconstructing an electrocardiogram signal from a photoplethysmogram based on deep learning according to claim 1, characterized in that, In step 2), the PPG signals and ECG signals are denoised by band-pass filtering, and the passband range of the filter is from 0.5 Hz to 40 Hz.

3. The method for reconstructing electrocardiogram signals from photoplethysmogram based on deep learning according to claim 1, wherein The generator in step 3) adopts a convolutional neural network combined with a long short-term memory network architecture, and its structure is: Wherein: CNN : the PPG signal processed by the convolutional neural network; LSTM( is the long short-term memory network module; : the parameters of the convolutional neural network; : the parameters of the long short-term memory network.

4. The method for reconstructing an electrocardiogram signal from a photoplethysmogram based on deep learning according to claim 3, wherein The output layer of the generator uses deconvolution operations to upsample the signal, and its formula is: Wherein: Intermediate feature representation of the generator; Deconv Transposed convolution operation; : Parameters of the transposed convolution operation.

5. The method for reconstructing electrocardiogram signals from photoplethysmogram based on deep learning according to claim 1, characterized in that, The supervised training in step 4) uses a multi-scale loss function, which is defined as: Wherein: : Total loss function; : Power spectral density difference loss in the frequency domain; Weight coefficient.

6. The method for reconstructing electrocardiogram signals from photoplethysmogram based on deep learning according to claim 5, wherein, Frequency domain loss It is defined as: Wherein: : The power spectral density of the real ECG signal; : The power spectral density of the generated ECG signal; Frequency.

7. The method for reconstructing electrocardiogram signals from photoplethysmogram based on deep learning as described in claim 1, wherein The model in step 6) is deployed on a smart bracelet, a smart watch or other wearable devices, and the real-time inference of the deep learning model is realized through embedded hardware.

8. The method for reconstructing electrocardiogram signals based on photoplethysmogram using deep learning as claimed in claim 1, wherein The sampling frequency of the PPG signals is from 100 Hz to 200 Hz, and the sampling frequency of the generated ECG signals is from 300 Hz to 500 Hz.

Images