A photoelectric plethysmogram signal repairing method based on segmented quality assessment and frequency domain constraint

Abstract

The application discloses a photoelectric plethysmogram signal repairing method based on segmented quality evaluation and frequency domain constraint, which comprises the following steps: preprocessing the obtained original photoelectric plethysmogram signal; performing power spectrum density analysis on the pretreated signal, and extracting the main frequency information in the frequency band corresponding to the heart rate; further dividing the pretreated signal into n non-overlapping signal segments, and performing quality scoring on each signal segment; generating a sine wave signal segment consistent with the length of the signal segment with the lowest quality score and the sampling rate according to the main frequency information, and replacing the signal segment with the lowest quality score to form a pseudo pulse wave signal; inputting the pseudo pulse wave signal into a signal repairing model for processing, and outputting a high-quality photoelectric plethysmogram signal. The application can repair abnormal segments in the continuous photoelectric plethysmogram signal, improve the continuity and stability of the signal, and help accurate analysis of the heart rate and related physiological parameters.

Description

Technical Field

[0001] This invention relates to the field of digital signal processing technology, and in particular to a method for repairing photoplethysmography (PPG) signals based on segmented quality assessment and frequency domain constraints. Background Technology

[0002] Photoplethysmography (PPG) for measuring heart rate and blood oxygen saturation has been widely used in clinical diagnosis and monitoring. Related portable and smart wearable medical products have also entered daily life, facilitating proactive health management, particularly playing a crucial role during periods of high respiratory infectious disease prevalence. The PPG signal measured by this technology can also be used to analyze parameters such as respiratory rate, blood pressure, and heart rate variability, which is essential for the effective prevention and timely diagnosis and treatment of cardiovascular diseases. It also has significant research value in areas such as emotion recognition. However, accurate analysis of these physiological parameters requires a continuous and stable PPG signal. Conventional PPG signals are measured using a contact-based single-point method. However, during the acquisition process, external factors such as the subject's movements can cause the photoelectric sensor to shift from the measured site, introducing noise interference that is often impossible to remove effectively using traditional methods such as filtering. This severely affects signal quality, leading to a decrease in the accuracy of subsequent physiological parameter estimation. As a further development of PPG technology, imaging photoplethysmography (IPPG) uses a camera as a light sensor to remotely record subtle changes in the intensity of reflected light from the skin surface caused by changes in subcutaneous blood volume, primarily in the face. This allows for non-contact, simultaneous measurement of multiple physiological parameters in natural settings, meeting the needs of some special scenarios. Furthermore, with the widespread adoption and performance improvements of high-definition cameras, this represents the future direction for physiological parameter monitoring equipment. Similar to contact PPG technology, IPPG technology faces the same challenges. Moreover, in non-contact measurements, changes in ambient light and movement can severely interfere with signal quality, leading to abnormal segments in the acquired photoplethysmogram, thus affecting the estimation of physiological parameters. Summary of the Invention

[0003] To address the problems existing in the background technology, the present invention aims to provide a method for repairing photoplethysmography (PPG) signals based on segmented quality assessment and frequency domain constraints. This method can effectively reduce the influence of abnormal segments in the source signal and improve the quality of PPG signals.

[0004] The specific technical solution for achieving the purpose of this invention is as follows:

[0005] (1) The acquired raw photoplethysmography pulse wave signal is preprocessed to obtain the preprocessed pulse wave signal. ;

[0006] The original photoplethysmography (PPG) signal obtained is a PPG signal acquired and extracted using contact or non-contact methods. The present invention does not limit the signal acquisition and extraction method.

[0007] The preprocessing includes two steps: first, bandpass filtering with a bandwidth of 0.5~3.5Hz is applied to the acquired raw photoplethysmography (PPG) signal; then, amplitude normalization is performed to obtain the processed PPG signal. .

[0008] (2) The preprocessed pulse wave signal Perform power spectral density analysis to extract the dominant frequency within the frequency band corresponding to heart rate. Specifically:

[0009] Calculate the preprocessed pulse wave signal Power spectral density in the frequency domain;

[0010] Within the preset heart rate frequency range, search for the maximum value in the power spectral density and take its corresponding frequency value as the dominant frequency. .

[0011] (3) The preprocessed pulse wave signal Divide into n non-overlapping signal segments { , … }, and score the quality of each signal segment to obtain the quality score sequence Q={ , … }; where n is a preset value;

[0012] The quality score refers to the preprocessed pulse wave signal. Input the quality scoring model to obtain the quality score for each signal segment. The quality scoring model can be the one described in invention patent ZL202411507519.9, or other photoplethysmography (PPG) signal quality assessment methods; this invention does not limit the specific methods used.

[0013] (4) Determine the lowest score in the quality score sequence. Corresponding signal segment This refers to the lowest signal segment in the quality score.

[0014] (5) Based on the aforementioned main frequency Generate the lowest signal segment of the quality score. Sine wave signal segments with consistent length and sampling rate ;

[0015] (6) The sine wave signal segment Replace the preprocessed pulse wave signal The lowest signal segment of the quality score mentioned in the text This generates a pseudo-pulse wave signal s'(t) = { , …, };

[0016] (7) The pseudo-pulse wave signal s'(t) is processed by the input signal repair model, and the repaired signal is output. .

[0017] The signal restoration model has an architecture that includes two sets of convolutional blocks, an embedding layer, a Transformer encoder layer, and a linear layer. Each set of convolutional blocks contains a one-dimensional convolutional layer (Conv1d), a batch normalization layer (BatchNorm1d), and an activation function (ReLU).

[0018] The training process of the signal restoration model is as follows:

[0019] Using the pseudo-pulse wave signal s'(t) as input and the standard pulse wave signal as label, the signal restoration model is obtained by using signal loss and gradient loss as a joint loss function to control training. The signal loss function is Pearson Loss, which compares the difference between the output signal of the signal restoration model and the label. The gradient loss function is Sobolev Loss, which compares the gradient similarity between the output signal of the signal restoration model and the label in the form of first derivatives, as shown in the following formula:

[0020] ,

[0021] ,

[0022] In the formula, N is the number of samples. It is the i-th signal value of a certain standard pulse wave signal. It is the mean value of a certain standard pulse wave signal. It is the i-th signal value of a certain repaired pulse wave signal. is the mean value of a certain segment of the pulse wave signal after repair, and m is the number of signal values ​​in a segment of the signal.

[0023] The beneficial effects of the technical solution provided by this invention are as follows: This invention proposes a method for repairing photoplethysmography (PPG) signals based on segmented quality assessment and frequency domain constraints. By segmenting and scoring the PPG signal to identify the low-quality signal segments requiring repair, and then combining the remaining high-quality segments with the dominant frequency information extracted through frequency domain analysis to generate sinusoidal wave segments to replace the signal segments requiring repair. Furthermore, a signal repair model ensures the continuity and stability of the signal. This method effectively optimizes the quality of abnormal signal segments and provides higher recovery accuracy. Using this method, PPG signals can be dynamically repaired, improving signal continuity and stability, thereby enhancing the accuracy and reliability of subsequent estimations of various physiological parameters. It has significant practical value and application prospects. Attached Figure Description

[0024] Figure 1 This is a flowchart of an embodiment of the present invention;

[0025] Figure 2 This is a schematic diagram of signal segmentation according to an embodiment of the present invention;

[0026] Figure 3 This is a schematic diagram of pseudo-pulse wave signal generation according to an embodiment of the present invention;

[0027] Figure 4 This is a diagram of the signal repair network structure according to an embodiment of the present invention. Detailed Implementation

[0028] To more clearly illustrate the technical means, technical improvements, and beneficial effects of this invention, the invention will be described in detail below with reference to the accompanying drawings.

[0029] Example

[0030] See Figure 1 This embodiment uses an IPPG signal with a duration of 4 seconds and a sampling rate of 30Hz as an example.

[0031] S101: Preprocess the acquired raw photoplethysmography (PPG) signal to obtain the processed PPG signal. .

[0032] The specific steps are as follows:

[0033] A raw photoplethysmography (PPG) signal was extracted from a 4-second face video (with a frame rate of 30 frames per second).

[0034] The acquired raw photoplethysmography (PPG) signal is bandpass filtered with a passband range of [0.5, 3.5] Hz;

[0035] The signal amplitude is normalized, i.e., the amplitude range is set between [0, 1], to obtain the preprocessed pulse wave signal. .

[0036] S102: The preprocessed pulse wave signal Perform power spectral density analysis to extract the dominant frequency within the frequency band corresponding to heart rate. .

[0037] Calculate the preprocessed pulse wave signal Power spectral density in the frequency domain;

[0038] Within the preset heart rate frequency range (0.5~3.5Hz in this embodiment), the maximum value in the power spectral density is searched, and the corresponding frequency value is taken as the dominant frequency. In this embodiment, =1.75Hz.

[0039] S103: The preprocessed pulse wave signal Divide into n non-overlapping signal segments { , …, }, and score the quality of each signal segment to obtain the quality score sequence Q={ , … }; where n is a preset value.

[0040] The specific steps are as follows:

[0041] The preprocessed pulse wave signal Divide the signal into 3 segments with a window width of 40 and a step size of 40. , , },like Figure 2 As shown.

[0042] The preprocessed pulse wave signal The input quality scoring model yields quality scores for three signal segments. , , This constitutes the mass fraction sequence Q={ , , In this example, Q = {0.75, 0.45, 0.81}.

[0043] In this embodiment, the quality scoring model is the quality scoring model described in invention patent ZL202411507519.9.

[0044] S104: Determine the lowest score in the quality score sequence. Corresponding signal segment This refers to the lowest signal segment in terms of quality score.

[0045] The specific steps are as follows:

[0046] Based on the quality score sequence Q={0.75, 0.45, 0.81}, the segment with the lowest quality score is the second segment. The lowest score is =0.45.

[0047] S105: Based on the aforementioned main frequency Generate the lowest signal segment of the quality score. Sine wave signal segments with consistent length and sampling rate .

[0048] The specific steps are as follows:

[0049] According to the main frequency described in this embodiment =1.75Hz, generating the lowest signal segment with the lowest quality score. Sine wave signal segments of uniform length Its amplitude is between [0, 1]. In this embodiment, the minimum signal segment length of the quality score is 40, and the sampling rate is 30Hz.

[0050] S106: The generated sine wave signal segment Replace the preprocessed pulse wave signal The lowest signal segment of the quality score mentioned in the text This generates a pseudo-pulse wave signal s'(t) = { , …, }

[0051] The specific steps are as follows:

[0052] The generated sine wave signal segment Replace the lowest signal segment of the quality score This generates a pseudo-pulse wave signal s' ={ , , },like Figure 3 As shown.

[0053] S107: Process the pseudo-pulse wave signal s'(t) input signal repair model and output the repaired signal. .

[0054] The specific steps are as follows:

[0055] The signal restoration model's architecture includes two sets of convolutional blocks, an embedding layer, a Transformer encoder layer, and a linear layer. Each convolutional block contains a one-dimensional convolutional layer (Conv1d), a batch normalization layer (BatchNorm1d), and a ReLU activation function (ReLU), as follows: Figure 4 As shown.

[0056] The training process of the signal restoration model is as follows:

[0057] Using the pseudo pulse wave signal s' Using standard pulse wave signals as input and standard pulse wave signals as labels, the signal restoration model is trained using a joint loss function of signal loss and gradient loss. The signal loss function is Pearson Loss, which compares the difference between the output signal of the signal restoration model and the label. The gradient loss function is Sobolev Loss, which compares the gradient similarity between the output signal of the signal restoration model and the label in the form of first derivatives, as shown in the following formula:

[0058] ,

[0059] ,

[0060] In the formula, N is the number of samples. It is the i-th signal value of a certain standard pulse wave signal. It is the mean value of a certain standard pulse wave signal. It is the i-th signal value of a certain repaired pulse wave signal. is the mean value of a certain segment of the pulse wave signal after repair, and m is the number of signal values ​​in a segment of the signal.

[0061] The pseudo pulse wave signal s'(t) is input into the trained signal restoration model, and the restored pulse wave signal is finally output. .

[0062] In summary, this invention proposes a photoplethysmography (PPG) signal restoration method based on segmented quality assessment and frequency domain constraints. This method accurately locates the signal segments requiring restoration by performing segmented quality scoring on the preprocessed PPG signal. Then, combined with frequency domain analysis, a sinusoidal signal segment with the same dominant frequency as the original signal is generated to initially replace the portion requiring restoration. Finally, a trained signal restoration network ensures the continuity and stability of the restored signal. This method can significantly improve the quality of PPG signals, especially those with abnormal quality, providing reliable technical support for the accurate measurement and analysis of subsequent physiological signals.

[0063] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for repairing photoplethysmography (PPG) signals based on segmented quality assessment and frequency domain constraints, characterized in that, The method includes the following specific steps: Step 1: Preprocess the acquired raw photoplethysmography (PPG) signal to obtain the preprocessed PPG signal. ; Step 2: Process the preprocessed pulse wave signal Perform power spectral density analysis to extract the dominant frequency within the frequency band corresponding to heart rate. ; Step 3: The preprocessed pulse wave signal Divide into n non-overlapping signal segments { , }, and score the quality of each signal segment to obtain the quality score sequence Q={ }; where n is a preset value; Step 4: Determine the lowest score in the quality score sequence. Corresponding signal segment This refers to the lowest signal segment in the quality score. Step 5: Based on the stated main frequency Generate a sinusoidal signal segment with the same length and sampling rate as the lowest signal segment in the quality score. ; Step 6: Extract the sine wave signal segment Replace the preprocessed pulse wave signal The lowest signal segment of the quality score mentioned in the text This generates a pseudo-pulse wave signal s' , ; Step 7: Convert the pseudo pulse wave signal s' The input signal is processed by the signal repair model, and the repaired signal is output. .

2. The photoplethysmography (PPG) signal restoration method according to claim 1, characterized in that, Step 1, the preprocessing described, includes two steps: first, bandpass filtering with a bandwidth of 0.5~3.5Hz is applied to the acquired raw photoplethysmography (PPG) signal; then, amplitude normalization is performed to obtain the preprocessed PPG signal. .

3. The photoplethysmography (PPG) signal restoration method according to claim 1, characterized in that, The specific process of step 2 is as follows: Calculate the preprocessed pulse wave signal Power spectral density in the frequency domain; Within the preset heart rate frequency range, search for the maximum value in the power spectral density and take its corresponding frequency value as the dominant frequency. .

4. The photoplethysmography (PPG) signal restoration method according to claim 1, characterized in that, The quality score mentioned in step 3 refers to the preprocessed pulse wave signal. Input the quality scoring model to obtain the quality score for each signal segment.

5. The photoplethysmography (PPG) signal restoration method according to claim 1, characterized in that, The signal restoration model described in step 7 has an architecture that includes two sets of convolutional blocks, an embedding layer, a Transformer encoder layer, and a linear layer; each set of convolutional blocks contains a one-dimensional convolutional layer, a batch normalization layer, and an activation function. The training process of the signal restoration model is as follows: Using the pseudo pulse wave signal s' Using standard pulse wave signals as input and standard pulse wave signals as labels, the signal loss and gradient loss are used as a joint loss function to control the training, resulting in the signal restoration model. The signal loss function is Pearson Loss, which compares the difference between the signal output by the signal restoration model and the label, and the gradient loss function is Sobolev Loss, which compares the gradient similarity between the signal output by the signal restoration model and the label in the form of the first derivative.

Images