Method and system for photoplethysmographic signal quality enhancement

Abstract

The application discloses a kind of photoelectric volume pulse wave signal quality enhancement method and system, the method includes the following steps: collecting PPG signal section to be processed and carrying out signal section quality assessment;For medium quality PPG signal section, extract its best quality continuous three pulse wave fragments, according to valley point index segmentation to obtain three pulse waves, interpolation to fixed length after taking the mean of three pulse waves, obtain reference pulse wave and extract its characteristic sequence, then input a predetermined fully connected neural network, generate PPG pulse wave template;Based on RR interval constraint and amplitude constraint, combined with PPG pulse wave template, reconstruct PPG signal section, carry out signal extension, boundary processing and signal smoothing processing, obtain quality enhanced PPG signal section.Using the technical scheme, on the basis of signal quality assessment, the quality of medium quality signal is enhanced, to improve the purpose of signal quality, improve the utilization rate of PPG signal and the accuracy of physiological index detection.

Description

Technical Field

[0001] This invention belongs to the field of signal processing technology and relates to a method and system for enhancing the quality of photoplethysmography (PPG) signals. Background Technology

[0002] Photoplethysmography (PPG) is a non-invasive optical detection technology used to measure changes in tissue blood volume. Currently, PPG is widely used in wearable devices such as smartwatches, rings, and health patches. Because these devices offer non-invasive detection, high real-time performance, ease of wear, and long-term monitoring, they can be used in daily life to extract heart rate and blood oxygen saturation in real time, assess heart rate variability, identify electrocardiographic rhythms, and measure blood pressure.

[0003] PPG signals are highly sensitive to motion interference, and motion interference is unavoidable in everyday wearable PPG signal detection devices. This results in low PPG signal quality, which seriously affects the accuracy of physiological indicators such as heart rate and blood oxygen saturation.

[0004] Recent studies have reported that wearable PPG signals, lasting from 44% to 86% of the duration, cannot be used to accurately extract heart rate metrics. Researchers have conducted extensive studies on motion interference suppression of PPG signals, but the improvement in signal quality is limited under complex motion interference conditions.

[0005] PPG signal quality enhancement is essentially a denoising technique. Traditional denoising methods use filters to remove interference and noise from signals in order to obtain high-quality signals. However, traditional filtering methods cannot effectively separate motion interference and photoplethysmography (PPG) signals, making it difficult to effectively improve filtering performance.

[0006] In existing technologies, methods based solely on signal quality assessment simply remove low-quality signals, resulting in data fragmentation and significant data loss. Some quality enhancement algorithms fail to reconstruct noisy PPG signals when noise levels are high. Furthermore, although deep learning-based models have shown acceptable performance on large datasets, they are limited to handling noise generated by specific activities (such as waving and handshaking), and they perform poorly in reconstructing key features such as dicrotic notches. Summary of the Invention

[0007] The purpose of this invention is to address the aforementioned problems in existing technologies by proposing a method and system for enhancing the quality of photoplethysmography (PPG) signals.

[0008] To achieve the above objectives, the basic solution of the present invention is: a method for enhancing the quality of photoplethysmography (PPG) signals, comprising the following steps:

[0009] The PPG signal segment to be processed is acquired, and the signal quality of the PPG signal segment to be processed is evaluated using the PPG signal segment quality assessment method to obtain the PPG signal segment quality classification result.

[0010] For medium-quality PPG signal segments, the three consecutive pulse wave segments with the best quality in the PPG signal segment are extracted using the local pulse wave segment quality assessment method.

[0011] Based on the valley point index in the pulse wave segment, the pulse wave segment is divided into three pulse waves. The three pulse waves are interpolated to a set length using an interpolation method. The mean of the three interpolated pulse waves is calculated to obtain the reference pulse wave.

[0012] Extract the characteristic parameter sequence of the reference pulse wave;

[0013] The extracted feature parameter sequence is input into a pre-defined fully connected neural network to generate a PPG pulse wave template.

[0014] Based on RR interval constraints and amplitude constraints, combined with PPG pulse wave templates, PPG signal segments are reconstructed.

[0015] The reconstructed PPG signal segment is subjected to signal extension, boundary processing, and signal smoothing to obtain a PPG signal segment with enhanced quality.

[0016] The working principle and beneficial effects of this basic scheme are as follows: This technical scheme enhances PPG signals based on feature-guided template reconstruction and physiological constraints. On the basis of signal quality assessment, it enhances the quality of signals with medium quality to improve signal quality, thereby improving the utilization rate of PPG signals and the accuracy of physiological indicator detection.

[0017] Furthermore, the method for evaluating the signal quality of the PPG signal segment to be processed using the PPG signal segment quality assessment method is as follows:

[0018] Before classification, the PPG signal is truncated into multiple signal segments of a certain duration. Then, after baseline drift removal, smoothing filtering and Zscore normalization preprocessing, the signal segments are evaluated for quality.

[0019] Fourteen feature parameters were extracted from four dimensions: PPG signal morphology, time-frequency spectrum, consistency and accuracy of detected physiological indicators, and information entropy. Specifically, in the morphological dimension, eight feature parameters were extracted, including pulse wave beat-by-beat similarity, pulse width similarity, amplitude similarity, and moving average similarity, based on peak and trough segmentation. In the time-frequency spectrum dimension, two feature parameters were extracted: the ratio of dominant frequency energy to total energy and the connectivity of the high-energy range of the time-frequency spectrum. In the dimension of consistency and accuracy of detected physiological indicators, two feature parameters were extracted: heart rate detection consistency and AC component detection accuracy. In the dimension of information entropy, two feature parameters were extracted: power spectral entropy and singular spectral entropy.

[0020] The 14 extracted feature parameters are input into the trained signal quality assessment model, which is based on the random forest machine learning algorithm and classifies the PPG signal segment into three levels: good, medium, and poor according to the input feature parameter sequence.

[0021] Based on the output of the signal quality assessment model, the PPG signal segment quality level is obtained.

[0022] It is simple to operate and easy to use.

[0023] Furthermore, for medium-quality PPG signal segments, the three consecutive pulse wave segments with the best quality in the PPG signal segment are extracted using a local pulse wave segment quality assessment method. The specific steps are as follows:

[0024] Extracting valley points of PPG signal segments;

[0025] Traversing all possible three consecutive pulse wave segments based on the valley point index, at least 4 valley points are needed to form 3 pulse waves;

[0026] For each candidate segment, extract the quality features of three pulse waves:

[0027] Similarity Index Each pulse wave is used as a template. Other pulse waves are dynamically time-normalized to obtain the similarity with the template pulse wave, resulting in a beat-by-beat similarity sequence. The maximum value is taken as the beat-by-beat similarity of the candidate segment, and then the reciprocal is taken as the beat-by-beat similarity index.

[0028] Pulse width similarity index : Calculate the interval between adjacent valley points within the candidate segment to obtain the pulse width sequence, and calculate the pulse width similarity index by subtracting the ratio of the standard deviation to the mean of the pulse width sequence from 1;

[0029] Amplitude Similarity Index The amplitude similarity index is calculated by subtracting the ratio of the standard deviation to the mean of the amplitude series from 1.

[0030] Moving average similarity index The pulse wave is divided according to the trough point index, the pulse waves are left-aligned, and the mean value is taken as the pulse wave moving average. The length of the moving average is the longest pulse wave length. Each pulse wave is correlated with the moving average of the same length to obtain the correlation coefficient. The average correlation coefficient is used as the moving average similarity, and then the reciprocal is taken to obtain the moving average similarity index.

[0031] The obtained pulse similarity index sequence, pulse width similarity index sequence, amplitude similarity index sequence, and moving average similarity index sequence are normalized to the [0,1] interval. The sequence length is consistent with the number of candidate segments, and the order of elements in the sequence is consistent with the order of candidate segments.

[0032] Calculate the overall quality score for each candidate segment:

[0033] ,

[0034] in The weights of the pulse similarity index, pulse width similarity index, amplitude similarity index, and moving average similarity index are respectively set to 0.4, 0.1, 0.3, and 0.2, respectively, based on their contribution to signal quality.

[0035] The three consecutive pulse wave segments with the highest overall amplitude quality score are considered to be of the best quality.

[0036] Using a local pulse wave segment quality assessment method, the three consecutive pulse wave segments with the best quality in the current signal segment are identified for subsequent use.

[0037] Furthermore, the method for extracting the characteristic parameter sequence of the reference pulse wave is as follows:

[0038] The reference pulse wave is extracted from the three consecutive pulse wave segments with the best quality. The extraction method is as follows:

[0039] Based on the valley point index of the three consecutive pulse wave segments with the best quality, three pulse waves are obtained. The length of the three pulse waves is fixed to a sequence of a set length by cubic spline interpolation. The mean pulse wave obtained by taking the average of the three sequences of the set length is the reference pulse wave.

[0040] For the reference pulse wave, extract 16 feature vectors in 4 dimensions, including:

[0041] Statistical characteristics: mean, standard deviation, root mean square, peak value, skewness, kurtosis;

[0042] Frequency domain characteristics: centroid frequency, mean square frequency, PSD entropy;

[0043] Waveform characteristics: waveform factor, peak factor, peak position ratio, peak rise ratio, first-order difference exponent;

[0044] Feature points: depth of diphtheria notch and ratio of diphtheria notch position.

[0045] From the best three consecutive pulse wave segments obtained, extract the feature vector of each complete pulse wave (from one trough to the next).

[0046] Furthermore, the method for generating a PPG pulse wave template by inputting the extracted feature parameter sequence of the reference pulse wave into a preset fully connected neural network model is as follows:

[0047] The fully connected neural network model includes one input layer, three cascaded hidden layers with "fully connected + LeakyReLU activation function" and one non-activated output layer; the input layer has a dimension of 16, corresponding to the 16 feature parameters of the pulse wave, each hidden layer has 128 fully connected neurons, and the output layer has a dimension of 50, corresponding to a pulse wave sequence of a set duration at a sampling rate of 50Hz.

[0048] The loss function of the fully connected neural network model uses a composite loss function, including mean squared error loss (MSE). ), first derivative loss ( Frequency domain loss () ) and loss of diabetic notch ( The total loss is:

[0049]

[0050] in These are the weights for the first derivative loss, frequency domain loss, and diabetic notch loss, respectively. The mean squared error loss is the basic loss function with a weight of 1.

[0051] The fully connected neural network model was trained using a high-quality dataset, and its parameters were tuned using a validation set to ultimately obtain... The values ​​are 0.6, 0.4, and 0.1 respectively.

[0052] After training the model, the PPG template generation network model is obtained;

[0053] The extracted pulse wave feature sequence is input into the model, and the model generates a single pulse wave sequence with a set duration, which is the PPG pulse wave template.

[0054] The pulse wave template serves as the reference for subsequent pulse wave generation, generating pulse waves of the same length based on the given pulse width.

[0055] Furthermore, based on RR interval constraints and amplitude constraints, and combined with the PPG pulse wave template, the PPG pulse wave signal segment is reconstructed. The specific steps are as follows:

[0056] RR interval constraint: Extract the valley points of the PPG signal segment, obtain the RR interval sequence of the PPG signal segment according to the valley point index, and obtain the reference RR interval according to the valley point index of the three consecutive pulse waves with the best quality.

[0057] The RR interval constraint is between 0.75 and 1.5 times the reference RR interval. An RR interval mean less than 0.75 times is considered too short and the corresponding valley point needs to be deleted; an RR interval mean greater than 1.5 times is considered too long and a valley point needs to be added between the two corresponding valley points.

[0058] After applying the RR interval constraint, the corrected valley point sequence is obtained;

[0059] The signal segment is divided into individual pulse waves based on the corrected valley point sequence. The pulse wave template is adapted to the length of the original pulse wave using cubic spline interpolation, and amplitude constraints are applied.

[0060] Amplitude constraint: The amplitude constraint range is set at 0.9-1.1 times the template amplitude. For pulse waves that exceed this range, they are limited to the amplitude range of 0.9-1.1 times by scaling.

[0061] The reconstructed pulse waves based on the template are connected sequentially to form a preliminary reconstructed PPG signal segment.

[0062] Achieving PPG signal quality enhancement based on physiological constraints.

[0063] Furthermore, the PPG signal segment after preliminary reconstruction is subjected to signal extension and boundary processing, specifically as follows:

[0064] For the left boundary, three template pulse waves are extended to the left based on the leftmost valley point; for the right boundary, three template pulse waves are extended to the right based on the rightmost valley point.

[0065] The signal segment is extracted based on the position information of the valley point at the starting point of the left extension and the valley point at the starting point of the right extension. That is, the leftmost and rightmost valley points before the extension are remembered, so that the length of the reconstructed signal is consistent with the length of the original signal, thus realizing the processing of the signal boundary.

[0066] By extending the signal, low-quality pulse waves near the boundary are processed.

[0067] Furthermore, the steps for smoothing the reconstructed PPG signal segment are as follows:

[0068] A smoothing filter is used to perform a 5-point smoothing filter on the reconstructed signal to eliminate spike interference, and then... Point smoothing filter, The baseline signal is obtained by using the signal sampling rate. The baseline signal is then subtracted from the signal after 5-point smoothing to eliminate baseline drift, thus obtaining the enhanced PPG signal segment.

[0069] The reconstructed signal has sharp points at the junctions of adjacent pulse waves and baseline inequality. Therefore, a smoothing filter is used to smooth the signal.

[0070] The present invention also provides a photoplethysmography (PPG) signal quality enhancement system, comprising a data acquisition unit and a processing unit, wherein the data acquisition unit is used to acquire the PPG signal segment to be processed and transmit it to the processing unit;

[0071] The processing unit executes the method described in this invention and outputs a PPG signal segment with enhanced quality.

[0072] This system utilizes a data acquisition unit and a processing unit to enhance the quality of signals with medium quality based on signal quality assessment, thereby improving signal quality. Attached Figure Description

[0073] Figure 1 This is a schematic flowchart of the photoplethysmography (PPG) signal quality enhancement method of the present invention;

[0074] Figure 2 This is a schematic diagram comparing the enhanced PPG signal and the original signal using the photoplethysmography (PPG) signal enhancement method of the present invention.

[0075] Figure 3 This is a comparative schematic diagram of the PPG signal quality enhancement method of the present invention—effective elimination of excessively short RR intervals;

[0076] Figure 4 This is a comparative schematic diagram of the PPG signal quality enhancement method of the present invention—effectively eliminating excessively long RR intervals;

[0077] Figure 5 This is a comparative schematic diagram of the PPG signal quality enhancement method of the present invention—effective processing of the left boundary (the valley point closest to 0 on the far left was not identified);

[0078] Figure 6 This is a comparative schematic diagram of the PPG signal quality enhancement method of the present invention—effective processing of the right boundary (the valley point closest to position 5 on the far right was not identified). Detailed Implementation

[0079] Embodiments of the present invention are described in detail below. Examples of these embodiments are shown in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and should not be construed as limiting the present invention.

[0080] In the description of this invention, it should be understood that the terms "longitudinal", "lateral", "up", "down", "front", "rear", "left", "right", "vertical", "horizontal", "top", "bottom", "inner", "outer", etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. They are only for the convenience of describing this invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on this invention.

[0081] In the description of this invention, unless otherwise specified and limited, it should be noted that the terms "installation", "connection" and "linking" should be interpreted broadly. For example, they can refer to mechanical or electrical connections, or internal connections between two components. They can be direct connections or indirect connections through an intermediate medium. Those skilled in the art can understand the specific meaning of the above terms according to the specific circumstances.

[0082] This invention discloses a method for enhancing the quality of photoplethysmography (PPG) signals. Based on signal quality assessment, it enhances signals with medium quality to improve signal quality, thereby increasing the utilization rate of PPG signals and the accuracy of physiological indicator detection. Figure 1 As shown, the method for enhancing the quality of photoplethysmography (PPG) signals includes the following steps:

[0083] Acquire the PPG signal segment to be processed, and use the PPG signal segment quality assessment method to assess the signal quality of the PPG signal segment to be processed (the assessment classifies the signal quality into three categories: good, medium, and poor, and applies signal quality enhancement to medium quality signals) to obtain the PPG signal segment quality classification results;

[0084] For medium-quality PPG signal segments (where signal quality is divided into three levels, poor-quality signals are usually difficult to distinguish individual pulse waves and are therefore generally discarded, while medium-quality signals can find three consecutive distinguishable pulse waves and are worth processing), the local pulse wave segment quality assessment method is used (the local pulse wave segment quality assessment method is relative to the full PPG signal segment quality assessment method; for a full 5-second signal segment, 14 feature parameters are extracted before quality assessment; for the quality assessment of three consecutive pulse waves in the signal segment, only 4 feature parameters are used, hence the name local pulse wave segment quality assessment method), to extract the three consecutive pulse wave segments with the best quality from the PPG signal segment;

[0085] Based on the valley point index in the pulse wave segment, the pulse wave segment is divided into three pulse waves. The three pulse waves are interpolated to a set length (e.g., 1 second) using an interpolation method. The mean of the three interpolated pulse waves is calculated to obtain the reference pulse wave.

[0086] Extract the characteristic parameter sequence of the reference pulse wave;

[0087] The extracted feature parameter sequence is input into a pre-defined fully connected neural network to generate a PPG pulse wave template.

[0088] Based on RR interval constraints and amplitude constraints, combined with PPG pulse wave templates, PPG signal segments are reconstructed.

[0089] The reconstructed PPG signal segment is subjected to signal extension, boundary processing, and signal smoothing to obtain a quality-enhanced PPG signal segment, such as... Figures 2-6 As shown.

[0090] In a preferred embodiment of the present invention, the method for evaluating the signal quality of the PPG signal segment to be processed using the PPG signal segment quality assessment method is as follows:

[0091] Before classification, the PPG signal is truncated into multiple signal segments of a certain duration. Then, after baseline drift removal, smoothing filtering and Zscore normalization preprocessing, the signal segments are evaluated for quality.

[0092] Fourteen feature parameters were extracted from four dimensions: PPG signal morphology, time-frequency spectrum, consistency and accuracy of detected physiological indicators, and information entropy. Specifically, in the morphological dimension, eight feature parameters were extracted, including pulse wave beat-by-beat similarity, pulse width similarity, amplitude similarity, and moving average similarity, based on peak and trough segmentation. In the time-frequency spectrum dimension, two feature parameters were extracted: the ratio of dominant frequency energy to total energy and the connectivity of the high-energy range of the time-frequency spectrum. In the dimension of consistency and accuracy of detected physiological indicators, two feature parameters were extracted: heart rate detection consistency and AC component detection accuracy. In the dimension of information entropy, two feature parameters were extracted: power spectral entropy and singular spectral entropy.

[0093] The methods for extracting each feature parameter are as follows:

[0094] First, the peak point coordinate sequence is obtained through the peak point and valley point identification program. Peak point amplitude sequence ( Valley point coordinate sequence ( ), Valley point amplitude sequence ( ), used for subsequent feature parameter identification.

[0095] Chase similarity ( Each pulse wave serves as a template, and other pulse waves are dynamically time-warped (DTW) to obtain their similarity to the template pulse wave. Pulse waves are segmented by peak and trough points.

[0096]

[0097] Similarity in beat analysis includes similarity based on peak points as segmentation points. Similarity with valley points as the dividing points .

[0098]

[0099]

[0100] Pulse width similarity ( Pulse width includes the interval between adjacent peak points and the interval between adjacent valley points. Pulse width similarity is calculated by subtracting the ratio of the standard deviation to the mean of the pulse width sequence from 1.

[0101]

[0102]

[0103]

[0104]

[0105] Amplitude similarity ( ): Amplitude includes peak amplitude and trough amplitude. Amplitude similarity is calculated by subtracting the ratio of the standard deviation to the mean of the amplitude sequence from 1.

[0106]

[0107]

[0108] Moving average similarity ( The pulse wave is divided into peak and trough points, the pulse waves are aligned to the left, and the mean value is taken as the pulse wave moving average. The length of the moving average is the longest pulse wave. Each pulse wave is correlated with the moving average of the same length to obtain the correlation coefficient. The average correlation coefficient is used as the moving average similarity.

[0109] Taking the peak point as an example, the steps for extracting the moving average similarity feature are as follows:

[0110] Receive pulse wave count , Refer to obtaining the sequence The number of elements;

[0111] Create a pulse wave moving average sequence L is the longest pulse wave length;

[0112] Iterate through each pulse wave, and assign the first pulse wave of each pulse wave to the first pulse wave. The elements (if they exist) are accumulated into the avr-th element. Each position is counted, and the result is divided by the cumulative count. The resulting value is then updated in the aVR array. This process is repeated until the traversal is complete, yielding the pulse wave moving average. ;

[0113] The similarity coefficient between each pulse wave and the pulse wave moving average is calculated using the following formula:

[0114]

[0115] Calculate the mean of all similarity coefficients to obtain the moving average similarity. .

[0116]

[0117] Similarly, we can obtain the moving average similarity based on the valley points:

[0118]

[0119] The ratio of main frequency energy to total energy ( ): Perform time-frequency analysis on the signal to obtain the time spectrum, and accumulate the frequency energies within the time spectrum to obtain the total energy. Determine the dominant frequency, perform spectral analysis on the signal, and identify the frequency corresponding to the maximum spectral amplitude as the dominant frequency. Determine the main frequency range, and set the main frequency range as ( , ) Hz; In the signal time-frequency spectrum, find the energy of frequencies falling within the dominant frequency range, and accumulate the found energy to obtain the dominant frequency energy. ;

[0120] Calculate the ratio of core frequency energy to total energy:

[0121]

[0122] High-energy range connectivity of the time spectrum ( ): Normalize the time spectrum and then binarize it: Set a threshold It is 0.95, for energies less than The frequency energy returns to zero, greater than or equal to The frequency energy is set to 1; the set of 1s in the binarized time spectrum is checked for 8-connectivity to obtain the number of 8-connected objects:

[0123]

[0124] Heart rate extraction consistency ( Heart rate (FFR) refers to the difference between the heart rate obtained using time-domain features and the heart rate obtained using frequency-domain features. It is calculated as follows: Perform an FFT on the signal to obtain a spectrum, and extract the maximum frequency. Heart rate obtained by calculating the frequency domain: ; Calculate the heart rate obtained in the time domain:

[0125] Peak heart rate:

[0126]

[0127] Heart rate obtained at the trough:

[0128]

[0129] Heart rate obtained in the time domain:

[0130]

[0131] Consistency in heart rate extraction calculation:

[0132]

[0133] PPG AC component extraction accuracy ( The AC component refers to the fluctuation amplitude of the PPG signal after baseline removal. The accuracy of AC component extraction is calculated as follows: Cubic spline interpolation is performed on the peak point sequence to obtain the peak point envelope sequence. Perform cubic spline interpolation on the valley point sequence to obtain the valley point envelope sequence. ;right and By subtracting the components, we obtain the communication component:

[0134]

[0135] Calculate the accuracy of AC component extraction:

[0136]

[0137] Power spectral entropy ( The information entropy of the signal power spectrum is calculated as follows: 1. Calculate the signal power spectrum; 2. Divide the power spectrum into groups and calculate the proportion of energy in each group to the total energy, thus obtaining the probability density function. ; the probability density function Substituting into the information entropy calculation formula, we obtain the power spectral entropy:

[0138] ,

[0139] Singular spectral entropy ( Using the singular value decomposition result as the object of information entropy calculation, the steps for calculating singular spectral entropy are as follows:

[0140] Original signal Process into a matrix :

[0141] ,

[0142] Let A be a signal sequence, and N be the signal length. Perform singular value decomposition on matrix A to obtain the singular value spectrum. ; Calculate the proportion of each singular value to the sum of all singular values ​​to obtain the probability density function. ; the probability density function Substituting into the information entropy calculation formula, we obtain the singular spectral entropy:

[0143] ,

[0144] The 14 extracted feature parameters are input into the trained signal quality assessment model. The signal quality assessment model is based on the random forest machine learning algorithm. According to the input feature parameter sequence, the PPG signal segment is divided into three levels: good, medium and poor. The specific level standard can be set in advance according to the value range of the feature parameters.

[0145] Based on the output of the signal quality assessment model, the PPG signal segment quality level is obtained.

[0146] In a preferred embodiment of the present invention, for a medium-quality PPG signal segment, a local pulse wave segment quality assessment method is used to extract the three consecutive pulse wave segments with the best quality from the PPG signal segment. The specific steps are as follows:

[0147] Extracting valley points of PPG signal segments;

[0148] Traversing all possible three consecutive pulse wave segments based on the valley point index, at least 4 valley points are needed to form 3 pulse waves;

[0149] For each candidate segment, extract the quality features of three pulse waves:

[0150] Similarity Index Each pulse wave is used as a template. Other pulse waves are dynamically time-normalized to obtain the similarity with the template pulse wave, resulting in a beat-by-beat similarity sequence. The maximum value is taken as the beat-by-beat similarity of the candidate segment, and then the reciprocal is taken as the beat-by-beat similarity index.

[0151] Pulse width similarity index : Calculate the interval between adjacent valley points within the candidate segment to obtain the pulse width sequence, and calculate the pulse width similarity index by subtracting the ratio of the standard deviation to the mean of the pulse width sequence from 1;

[0152] Amplitude Similarity Index The amplitude similarity index is calculated by subtracting the ratio of the standard deviation to the mean of the amplitude series from 1.

[0153] Moving average similarity index The pulse wave is segmented according to the trough index, left-aligned, and the mean value is taken as the pulse wave moving average. The moving average length is the longest pulse wave length. Each pulse wave is correlated with the moving average of the same length to obtain the correlation coefficient. The average correlation coefficient is used as the moving average similarity, and the reciprocal is taken to obtain the moving average similarity index. The calculation methods for amplitude similarity index and moving average similarity index are the same as those for amplitude similarity and moving average similarity in signal segment quality assessment, only the application objects are different. The former is used for signal segments, and the latter is used for three pulse wave segments. Here, they are distinguished by parameter names.

[0154] The obtained pulse similarity index sequence, pulse width similarity index sequence, amplitude similarity index sequence, and moving average similarity index sequence are normalized to the [0,1] interval. The sequence length is consistent with the number of candidate segments, and the order of elements in the sequence is consistent with the order of candidate segments.

[0155] Calculate the overall quality score for each candidate segment:

[0156]

[0157] in The weights of the pulse similarity index, pulse width similarity index, amplitude similarity index, and moving average similarity index are respectively set to 0.4, 0.1, 0.3, and 0.2, respectively, based on their contribution to signal quality.

[0158] The three consecutive pulse wave segments with the highest overall amplitude quality score are considered to be of the best quality.

[0159] In a preferred embodiment of the present invention, the method for extracting the characteristic parameter sequence of the reference pulse wave is as follows:

[0160] The reference pulse wave is extracted from the three consecutive pulse wave segments with the best quality. The extraction method is as follows:

[0161] Based on the valley point index of the three consecutive pulse wave segments with the best quality, three pulse waves are obtained. By resampling or cubic spline interpolation, the length of the three pulse waves is fixed to a 1-second sequence. The mean pulse wave obtained by taking the average of the three 1-second sequences is the reference pulse wave.

[0162] For the reference pulse wave, extract 16 feature vectors in 4 dimensions, including:

[0163] Statistical characteristics: mean, standard deviation, root mean square, peak value, skewness, kurtosis;

[0164] Frequency domain characteristics: centroid frequency, mean square frequency, PSD entropy;

[0165] Waveform characteristics: waveform factor, peak factor, peak position ratio, peak rise ratio, first-order difference exponent;

[0166] Feature points: depth of diphtheria notch and ratio of diphtheria notch position.

[0167] In a preferred embodiment of the present invention, the method for generating a PPG pulse wave template by inputting the extracted feature parameter sequence of the reference pulse wave into a preset fully connected neural network model is as follows:

[0168] The fully connected neural network model consists of one input layer, three cascaded hidden layers with "fully connected + LeakyReLU activation function" and one non-activated output layer; the input layer has a dimension of 16, corresponding to the 16 feature parameter sequences of the pulse wave, each hidden layer has 128 fully connected neurons, and the output layer has a dimension of 50, corresponding to the pulse wave sequence of a set duration (e.g., 1 second) at a sampling rate of 50Hz.

[0169] The loss function of the fully connected neural network model uses a composite loss function, including mean squared error loss (MSE). ), first derivative loss ( Frequency domain loss () ) and loss of diabetic notch ( The total loss is:

[0170] ,

[0171] in These are the weights for the first derivative loss, frequency domain loss, and diabetic notch loss, respectively. The mean squared error loss is the basic loss function with a weight of 1.

[0172] The fully connected neural network model was trained using a high-quality dataset, and its parameters were tuned using a validation set to ultimately obtain... The values ​​are 0.6, 0.4, and 0.1 respectively.

[0173] After training the model, the PPG template generation network model is obtained;

[0174] The extracted pulse wave feature sequence is input into the model, and the model generates a pulse wave sequence with a duration of 1 second, which is the PPG pulse wave template. The pulse wave template serves as the reference for subsequent pulse wave generation, and pulse waves of the same length are generated according to the given pulse width.

[0175] This invention generates a standard pulse wave (the morphology of which is determined by the three best consecutive pulse waves in that signal segment). Subsequently, pulse waves of different lengths are generated based on different RR intervals; although the lengths differ, the morphology remains the same as the template pulse wave. This invention does not directly use the average of the three best pulse waves as the template because even the three best consecutive pulse waves in a medium-quality signal segment have a worse waveform than a high-quality pulse wave and cannot be used directly. Therefore, a model needs to be trained using a high-quality pulse wave dataset to achieve the goal of generating a high-quality template based on feature parameters.

[0176] In a preferred embodiment of the present invention, the PPG pulse wave signal segment is reconstructed based on RR interval constraints and amplitude constraints, combined with a PPG pulse wave template. The specific steps are as follows:

[0177] RR interval constraint: Extract the valley points of the PPG signal segment, obtain the RR interval sequence of the PPG signal segment according to the valley point index, and obtain the reference RR interval according to the valley point index of the three consecutive pulse waves with the best quality.

[0178] The RR interval constraint is between 0.75 and 1.5 times the reference RR interval. An RR interval mean less than 0.75 times is considered too short and the corresponding valley point needs to be deleted; an RR interval mean greater than 1.5 times is considered too long and a valley point needs to be added between the two corresponding valley points.

[0179] After applying the RR interval constraint, the corrected valley point sequence is obtained;

[0180] The signal segment is divided into individual pulse waves based on the corrected valley point sequence. The pulse wave template is adapted to the length of the original pulse wave using cubic spline interpolation, and amplitude constraints are applied.

[0181] Amplitude constraint: The amplitude constraint range is set at 0.9-1.1 times the template amplitude. For pulse waves that exceed this range, they are limited to the amplitude range of 0.9-1.1 times by scaling.

[0182] The reconstructed pulse waves based on the template are connected sequentially to form a preliminary reconstructed PPG signal segment.

[0183] In a preferred embodiment of the present invention, the purpose of signal extension is to process low-quality pulse waves near the boundary. Specifically, the pre-reconstructed PPG signal segment undergoes signal extension and boundary processing, which includes:

[0184] For the left boundary, three template pulse waves are extended to the left based on the leftmost valley point; for the right boundary, three template pulse waves are extended to the right based on the rightmost valley point.

[0185] After the extension is completed, the length of the signal segment will be greater than the length of the original signal. The signal segment is truncated based on the position information of the valley point of the extension start point on the left and the valley point of the extension start point on the right. That is, the leftmost and rightmost valley points before the extension are remembered, so that the length of the reconstructed signal is consistent with the length of the original signal, thus realizing the processing of the signal boundary.

[0186] In a preferred embodiment of the present invention, the step of performing signal smoothing processing on the reconstructed PPG signal segment is as follows:

[0187] The reconstructed signal exhibits sharp peaks at the junctions of adjacent pulse waves and baseline inequality. A smoothing filter is applied to the reconstructed signal using a 5-point smoothing filter to eliminate peak interference, followed by further processing. Point smoothing filter, The baseline signal is obtained by using the signal sampling rate. Subtracting the baseline signal from the 5-point smoothed signal eliminates baseline drift, resulting in the enhanced PPG signal segment. The input parameter of the smoothing filter is the number of points. 5-point smoothing takes four points forward from the current filtering point, and the average of these five points is used as the filtered value for the current point. Conversely, 2f+1-point smoothing produces the baseline signal; subtracting the baseline signal from the 5-point smoothed signal eliminates baseline drift.

[0188] This invention also provides a photoplethysmography (PPG) signal quality enhancement system, comprising a data acquisition unit and a processing unit. The data acquisition unit acquires PPG signal segments to be processed and transmits them to the processing unit. The processing unit executes the method described in this invention and outputs enhanced PPG signal segments.

[0189] Compared to traditional filtering methods, this invention is better able to identify the source of interference and then process it accordingly based on its severity. Instead of directly discarding poor-quality signals, the algorithm can retain signals of moderate quality and is also effective for some poor-quality signals, improving signal utilization and the accuracy of physiological indicator detection. Compared to deep learning algorithms, this invention has clearer and simpler steps and is more likely to preserve physiological features such as diphtheria notches.

[0190] The specific embodiments described herein are merely illustrative examples of the present invention. Those skilled in the art can make various modifications or additions to the described specific embodiments or use similar methods to replace them, without departing from the technology of the present invention or exceeding the scope defined by the appended claims.

Claims

1. A method for enhancing the quality of photoplethysmography (PPG) signals, characterized in that, Includes the following steps: The PPG signal segment to be processed is acquired, and the signal quality of the PPG signal segment to be processed is evaluated using the PPG signal segment quality assessment method to obtain the PPG signal segment quality classification results, which are divided into three levels: good, medium and poor. For medium-quality PPG signal segments, the three consecutive pulse wave segments with the best quality in the PPG signal segment are extracted using the local pulse wave segment quality assessment method. Based on the valley point index of the three consecutive pulse wave segments with the best quality, the pulse wave segment is divided into three pulse waves. The three pulse waves are interpolated to a set length using an interpolation method. The mean of the three interpolated pulse waves is calculated to obtain the reference pulse wave. Extract the characteristic parameter sequence of the reference pulse wave; The extracted feature parameter sequence is input into a pre-defined fully connected neural network to generate a PPG pulse wave template. Based on RR interval constraints and amplitude constraints, combined with PPG pulse wave templates, PPG signal segments are reconstructed. The reconstructed PPG signal segment is subjected to signal extension, boundary processing, and signal smoothing to obtain a PPG signal segment with enhanced quality.

2. The method for enhancing the quality of photoplethysmography (PPG) signals according to claim 1, characterized in that, The method for evaluating the signal quality of the PPG signal segment to be processed using the PPG signal segment quality assessment method is as follows: Before classification, the PPG signal is truncated into multiple signal segments of a certain duration. Then, after baseline drift removal, smoothing filtering and Zscore normalization preprocessing, the signal segments are evaluated for quality. Fourteen feature parameters were extracted from four dimensions: PPG signal morphology, time spectrum, consistency and accuracy of detected physiological indicators, and information entropy. In the morphological dimension, eight feature parameters were extracted from two groups: pulse wave beat-by-beat similarity, pulse width similarity, amplitude similarity, and moving average similarity, based on peak point segmentation and valley point segmentation. In the time-frequency spectrum dimension, two feature parameters are extracted: the ratio of dominant frequency energy to total energy and the connectivity of the high-energy range of the time-frequency spectrum. In terms of the consistency and accuracy of physiological indicators, two feature parameters were extracted: consistency of heart rate detection and accuracy of AC component detection. In terms of information entropy, two feature parameters are extracted: power spectral entropy and singular spectral entropy. The 14 extracted feature parameters are input into the trained signal quality assessment model, which is based on the random forest machine learning algorithm and classifies the PPG signal segment into three levels: good, medium, and poor according to the input feature parameter sequence. Based on the output of the signal quality assessment model, the PPG signal segment quality level is obtained.

3. The method for enhancing the quality of photoplethysmography (PPG) signals according to claim 2, characterized in that, For medium-quality PPG signal segments, the three consecutive pulse wave segments with the best quality in the PPG signal segment are extracted using a local pulse wave segment quality assessment method. The specific steps are as follows: Extracting valley points of PPG signal segments; Traversing all possible three consecutive pulse wave segments based on the valley point index, at least 4 valley points are needed to form 3 pulse waves; For each candidate segment, extract the quality features of four pulse waves: Similarity Index Each pulse wave is used as a template. Other pulse waves are dynamically time-normalized to obtain the similarity with the template pulse wave, resulting in a beat-by-beat similarity sequence. The maximum value is taken as the beat-by-beat similarity of the candidate segment, and then the reciprocal is taken as the beat-by-beat similarity index. Pulse width similarity index : Calculate the interval between adjacent valley points within the candidate segment to obtain the pulse width sequence, and calculate the pulse width similarity index by subtracting the ratio of the standard deviation to the mean of the pulse width sequence from 1; Amplitude Similarity Index The amplitude similarity index is calculated by subtracting the ratio of the standard deviation to the mean of the trough amplitude sequence from 1. Moving average similarity index The pulse wave is divided according to the valley point index, the pulse waves are left-aligned, and the mean value is taken as the pulse wave moving average. The length of the moving average is the longest pulse wave. Each pulse wave is correlated with the moving average of the same length to obtain the correlation coefficient. The average correlation coefficient is used as the moving average similarity, and then the reciprocal is taken to obtain the moving average similarity index. The obtained pulse similarity index sequence, pulse width similarity index sequence, amplitude similarity index sequence, and moving average similarity index sequence are normalized to the [0,1] interval. The sequence length is consistent with the number of candidate segments, and the order of elements in the sequence is consistent with the order of candidate segments. Calculate the overall quality score for each candidate segment: ， in The weights of the pulse similarity index, pulse width similarity index, amplitude similarity index, and moving average similarity index are respectively set to 0.4, 0.1, 0.3, and 0.2, respectively, based on their contribution to signal quality. The three consecutive pulse wave segments with the highest overall amplitude quality score are considered to be of the best quality.

4. The method for enhancing the quality of photoplethysmography (PPG) signals according to claim 3, characterized in that, The method for extracting the characteristic parameter sequence of the reference pulse wave is as follows: The reference pulse wave is extracted from the three consecutive pulse wave segments with the best quality. The extraction method is as follows: Based on the valley point index of the three consecutive pulse wave segments with the best quality, three pulse waves are obtained. The length of the three pulse waves is fixed to a sequence of a set length by cubic spline interpolation. The mean pulse wave obtained by taking the average of the three sequences of the set length is the reference pulse wave. For the reference pulse wave, extract 16 feature vectors in 4 dimensions, including: Statistical characteristics: mean, standard deviation, root mean square, peak value, skewness, kurtosis; Frequency domain characteristics: centroid frequency, mean square frequency, PSD entropy; Waveform characteristics: waveform factor, peak factor, peak position ratio, peak rise ratio, first-order difference exponent; Feature points: depth of diphtheria notch and ratio of diphtheria notch position.

5. The method for enhancing the quality of photoplethysmography (PPG) signals according to claim 1, characterized in that, The method for generating a PPG pulse wave template by inputting the extracted feature parameter sequence of the reference pulse wave into a preset fully connected neural network model is as follows: The fully connected neural network model includes one input layer, three cascaded hidden layers with "fully connected + LeakyReLU activation function" and one non-activated output layer; the input layer has a dimension of 16, corresponding to the 16 feature parameters of the pulse wave, each hidden layer has 128 fully connected neurons, and the output layer has a dimension of 50, corresponding to a pulse wave sequence of a set length at a sampling rate of 50Hz. The loss function of the fully connected neural network model uses a composite loss function, including mean squared error loss. First derivative loss Frequency domain loss and loss of diabetic notch The total loss is: ， in These are the weights for the first derivative loss, frequency domain loss, and diabetic notch loss, respectively. The mean squared error loss is the basic loss function with a weight of 1. The fully connected neural network model was trained using a high-quality dataset, and its parameters were tuned using a validation set to ultimately obtain... The values ​​are 0.6, 0.4, and 0.1 respectively. After training the model, the PPG template generation network model is obtained; The extracted pulse wave feature sequence is input into the model, and the model generates a pulse wave sequence of a set length, which is the PPG pulse wave template.

6. The method for enhancing the quality of photoplethysmography (PPG) signals according to claim 1, characterized in that, Based on RR interval constraints and amplitude constraints, and combined with the PPG pulse wave template, the PPG pulse wave signal segment is reconstructed. The specific steps are as follows: RR interval constraint: Extract the valley points of the PPG signal segment, obtain the RR interval sequence of the PPG signal segment according to the valley point index, and obtain the reference RR interval according to the valley point index of the three consecutive pulse waves with the best quality. The RR interval constraint is between 0.75 and 1.5 times the reference RR interval. An RR interval mean less than 0.75 times is considered too short and the corresponding valley point needs to be deleted; an RR interval mean greater than 1.5 times is considered too long and a valley point needs to be added between the two corresponding valley points. After applying the RR interval constraint, the corrected valley point sequence is obtained; The signal segment is divided into individual pulse waves based on the corrected valley point sequence. The pulse wave template is adapted to the length of the original pulse wave using cubic spline interpolation, and amplitude constraints are applied. Amplitude constraint: The amplitude constraint range is set at 0.9-1.1 times the template amplitude. For pulse waves that exceed this range, they are limited to the amplitude range of 0.9-1.1 times by scaling. The reconstructed pulse waves based on the template are connected sequentially to form a preliminary reconstructed PPG signal segment.

7. The method for enhancing the quality of photoplethysmography (PPG) signals according to claim 1, characterized in that, The PPG signal segment after preliminary reconstruction is subjected to signal extension and boundary processing, specifically as follows: For the left boundary, three template pulse waves are extended to the left based on the leftmost valley point; for the right boundary, three template pulse waves are extended to the right based on the rightmost valley point. The signal segment is extracted based on the position information of the valley point at the starting point of the left extension and the valley point at the starting point of the right extension. That is, the leftmost and rightmost valley points before the extension are remembered, so that the length of the reconstructed signal is consistent with the length of the original signal, thus realizing the processing of the signal boundary.

8. The method for enhancing the quality of photoplethysmography (PPG) signals according to claim 7, characterized in that, The steps for smoothing the reconstructed PPG signal segment are as follows: A smoothing filter is used to perform a 5-point smoothing filter on the reconstructed signal to eliminate spike interference, and then... Point smoothing filter, The baseline signal is obtained by using the signal sampling rate. The baseline signal is then subtracted from the signal after 5-point smoothing to eliminate baseline drift, thus obtaining the enhanced PPG signal segment.

9. A photoplethysmography (PPG) signal quality enhancement system, characterized in that, It includes a data acquisition unit and a processing unit. The data acquisition unit is used to acquire the PPG signal segment to be processed and transmit it to the processing unit. The processing unit executes the method described in any one of claims 1-8 and outputs a quality-enhanced PPG signal segment.

Images