A blood oxygen saturation estimation method and system based on PPG signal

Abstract

The present invention discloses a blood oxygen saturation estimation method and system based on PPG signals, which relate to the technical field of blood oxygen saturation estimation and include the following steps: S1: designing a bandpass filter based on a Butterworth filter, designing a comb filter based on the characteristics of motion artifacts in PPG signals, and removing motion artifacts in the PPG signals by using the bandpass filter and the comb filter; S2: performing preliminary screening of PGG signal features by a threshold filtering method based on a correlation coefficient; S3: performing wrapping method feature screening from three dimensions of time domain, frequency domain and statistics by a recursive feature elimination method, and extracting a PGG signal feature set; S4: estimating the PGG signal feature set by an integrated regressor based on a random forest algorithm; and being able to extract blood oxygen saturation information from human PPG signals and realize real-time blood oxygen saturation monitoring.

Description

Technical Field

[0001] The present invention relates to the technical field of blood oxygen saturation estimation, and in particular to a blood oxygen saturation estimation method and system based on PPG signals. Background Art

[0002] In intensive care, blood oxygen saturation is one of the important monitoring indicators of the patient's physiological status. Blood oxygen saturation refers to the proportion of hemoglobin bound to oxygen in the blood to the total hemoglobin. It is an important indicator for measuring the function of the respiratory and circulatory systems.

[0003] Existing blood oxygen saturation prediction methods based on PPG signals mainly calculate blood oxygen saturation by using the relationship between the ratio of the AC and DC amplitude ratios of the two received light signals and blood oxygen saturation. However, the accuracy of this method is still subject to many challenges. The empirical relationship curve between the modulation ratio and blood oxygen saturation is obtained from calibration data, and the quality of the calibration data directly affects the accuracy of the blood oxygen saturation measurement value. In addition, the empirical relationship curve between the modulation ratio and blood oxygen saturation is also affected by factors such as the measurement equipment and the measurement environment. The optimal relationship curve between the modulation ratio and blood oxygen saturation is different for different equipment or different test subjects. When the performance of the equipment changes or the physiological state of the test subject changes, the blood oxygen saturation measurement value lacks accuracy. At the same time, existing blood oxygen saturation measurement instruments are also easily affected by noise and motion artifacts. The measurement process requires the patient to remain still during measurement.

[0004] Therefore, a blood oxygen saturation estimation method and system based on PPG signal are provided to solve the above problems. Summary of the Invention

[0005] The purpose of the present invention is to provide a blood oxygen saturation estimation method and system based on PPG signals, which can monitor the dynamic changes of blood oxygen saturation in real time, improve the generalization ability and prediction accuracy of the prediction model, and provide more accurate blood oxygen saturation estimation.

[0006] To achieve the above objectives, the present invention provides a method and system for estimating blood oxygen saturation based on PPG signals, comprising the following steps:

[0007] S1: Design a bandpass filter based on the Butterworth filter, and design a comb filter based on the characteristics of motion artifacts in the PGG signal. Use the bandpass filter and comb filter to remove motion artifacts and noise in the PGG signal.

[0008] S2: Extract PGG signal features and perform preliminary screening of PGG signal features using a threshold filtering method based on correlation coefficient;

[0009] S3: Use recursive feature elimination to perform parcel feature screening from the time domain, frequency domain and statistical dimensions to extract the PGG signal feature set X selected ;

[0010] S4: The PGG signal feature set X is regressed by an integrated regressor based on the random forest algorithm. selected Make an estimate.

[0011] Preferably, in step S1, designing a bandpass filter based on a Butterworth filter specifically includes the following steps:

[0012] Step 1: Filter out the signals below the heart rate ω through a high-pass filter HR PPG signal components, the order of the high-pass filter is n h Set to 2, the cutoff frequency of the high-pass filter ω ch Set to 0.5ω HR , high-pass filter at heart rate ω HR The gain at the heart rate is 97.0%, and the high-pass filter is at the heart rate ω HR When the gain is greater than 95%, the order n h With the cutoff frequency ω ch Satisfy between:

[0013]

[0014] Among them H c (jω HR ) represents the high-pass filter at heart rate ω HR Gain at

[0015] Step 2: Filter out high-frequency noise through a low-pass filter. The order of the low-pass filter is n. l Set to 2, the cutoff frequency of the low-pass filter ω ch When set to 10Hz, the frequency of the effective information of the PPG signal is lower than 8Hz, and the gain of the low-pass filter at 8Hz is 84.2%;

[0016] Step 3: Connect the high-pass filter and the low-pass filter in series to obtain a band-pass filter. The transfer function of the band-pass filter is set to:

[0017]

[0018] Among them H bp (jω) represents the frequency response function of the bandpass filter, H hp (jω) represents the frequency response function of the high-pass filter, H lp (jω) represents the frequency response function of the low-pass filter.

[0019] Preferably, in step S1, the base frequency of the comb filter is set to the heart rate ωHR , estimate the heart rate ω through frequency domain analysis algorithm HR , specifically including the following steps:

[0020] Step 1: Convert the original PPG signal X and the original sampling frequency and heart rate estimation range Input frequency domain analysis algorithm, heart rate estimation range Determined through experience;

[0021] Step 2: Perform FFT transformation on the original PPG signal X to obtain the heart rate estimation range The signal amplitude S at each frequency ω within x ;

[0022] Step 3: Calculate the average amplitude of the original PPG signal X at the frequency ω And the heart rate ω HR Set to the average amplitude of the octave The maximum value of the frequency ω, the average amplitude of the frequency doubling Set to:

[0023]

[0024] Where n represents an integer rounded down, a kω Indicates the gain at the kth frequency.

[0025] Preferably, step S2 specifically includes the following steps:

[0026] S21: Calculate the Spearman rank correlation coefficient between the PPG signal and blood oxygen saturation;

[0027] S22: Preliminary screening of PPG signal features using the Spearman rank correlation coefficient to remove signals below the threshold r threshold The Spearman rank correlation coefficient corresponds to the PPG signal characteristics, the threshold r threshold Set to 0.1, Spearman rank correlation coefficient r s Set to:

[0028]

[0029] R(X) and R(Y) both represent random variable samples, σ R(X) and σ R(Y) Both represent the standard deviation of the random variable sample.

[0030] Preferably, step S3 specifically includes the following steps:

[0031] S31: Determine a base model, which is set to a linear regression model, a support vector machine model, or a random forest model;

[0032] S32: training the base model using PGG signal features;

[0033] S33: calculating the importance scores I of different PGG signal features using the base model, and ranking the importance scores I;

[0034] S34: Eliminate the PGG signal feature with the lowest importance score I;

[0035] S35: Retrain the base model using the PGG signal features after elimination. If the training result meets the performance evaluation index or feature quantity standard, extract the PGG signal feature set X corresponding to the training result. selected , otherwise repeat steps S31-S34.

[0036] Preferably, in step S35, the number of features is set to 6, and the 6 features are respectively the average value of the systolic peak height, the duration of the diastolic peak at 33% height, the signal average value, the signal upper quartile value, the power P of 0-2Hz, and the average value of the systolic peak height. [0-2Hz] , the ratio of the signal standard deviation to the median value r sd / m ;

[0037] 0-2Hz power P [0-2Hz] Set to:

[0038]

[0039] Where X(f) represents the Fourier transform of the signal;

[0040] The ratio of the signal standard deviation to the median value r sd / m Set to:

[0041]

[0042] where x i represents a single sample value, μ represents the mean value of the signal, and median represents the median value of the signal.

[0043] Preferably, step S4 specifically includes the following steps:

[0044] S41: Train N decision trees using N random subsets of the sample set;

[0045] S42: Recursively split the node, randomly select n f node features, and split the optimal features and optimal thresholds. The splitting stop condition is set to the node depth or the number of node samples. The feature with the smallest mean square error after splitting is selected as the optimal feature of the node.

[0046] S43: Record the optimal features and thresholds of all nodes to form a decision tree. N decision trees form a random forest.

[0047] S44: Starting from the root node of the decision tree, the decision is made based on the node's features and thresholds, and the node is transferred to the left or right child node. If the node is transferred to a leaf node, the estimated value of the leaf node is recorded.

[0048] S45: Calculate the average of the estimated values ​​of the leaf nodes of N decision trees to obtain the final estimated value Final estimate Set to:

[0049]

[0050] Where B represents the number of decision trees, and the number of decision trees B is set to 1000. b (x) represents the final estimated value of the bth decision tree;

[0051] S46: Combine performance evaluation indicators to evaluate the consistency between the estimated value and the true value, and use the support vector machine algorithm, decision tree algorithm and multi-layer perceptron algorithm for comparative testing.

[0052] Preferably, in step S46, the performance evaluation indicators include mean absolute error (MAE) and mean square error (MSE), and the calculation methods of mean absolute error (MAE) and mean square error (MSE) are respectively:

[0053]

[0054] N s represents the total number of samples, y j Indicates the predicted value of blood oxygen, Indicates the actual value of blood oxygen.

[0055] Preferably, in step S46, the consistency between the estimated value and the true value is evaluated by a Bland-Altman plot, the horizontal axis of the Bland-Altman plot is set to the average of the predicted value and the estimated value, and the vertical axis of the Bland-Altman plot is set to the difference between the estimated value and the true value.

[0056] Preferably, a system for a blood oxygen saturation estimation method based on a PPG signal includes a motion artifact removal module, a signal feature extraction module and a blood oxygen saturation estimation module.

[0057] Therefore, the present invention adopts the above-mentioned blood oxygen saturation estimation method and system based on PPG signal, which has the following beneficial effects:

[0058] (1) Design a bandpass filter to remove motion artifacts and noise to prevent the patient's motion from affecting blood oxygen saturation monitoring and ensure the accuracy and stability of the measurement results;

[0059] (2) PPG signals can provide real-time blood oxygen saturation data, which can be continuously monitored for a long time and record dynamic changes, helping to identify potential health problems;

[0060] (3) By integrating multiple decision trees through random forest, overfitting can be effectively reduced, the generalization ability and prediction accuracy of the model can be improved, the complex nonlinear relationship between the waveform characteristics and frequency components in the PPG signal can be captured, and a more accurate blood oxygen saturation estimation can be provided.

[0061] The method scheme of the present invention is further described in detail below through the drawings and examples. BRIEF DESCRIPTION OF THE DRAWINGS

[0062] Figure 1 Flowchart of the feature screening algorithm of the present invention;

[0063] Figure 2 Schematic diagram of MAE and MSE for different feature numbers in the present invention;

[0064] Figure 3 is the Bland-Altman graph of the random forest algorithm of the present invention;

[0065] Figure 4 is a Bland-Altman diagram of the support vector machine algorithm used in the comparative test of the present invention;

[0066] Figure 5 is the Bland-Altman diagram of the decision tree algorithm in the comparative test of the present invention;

[0067] Figure 6 This is the Bland-Altman graph of the 100-node multilayer perceptron algorithm used in the comparative test of the present invention;

[0068] Figure 7 This is the Bland-Altman graph of the 400-node multilayer perceptron algorithm used in the comparative test of the present invention. DETAILED DESCRIPTION

[0069] The method scheme of the present invention is further described below through the drawings and examples.

[0070] Unless otherwise defined, technical terms or scientific terms used in the present invention shall have the same meaning as commonly understood by one of ordinary skill in the art to which the present invention belongs.

[0071] The words “include” or “comprising” and similar words used in the present invention mean that the elements before the word include the elements listed after the word, and do not exclude the possibility of also including other elements. The orientation or position relationship indicated by the terms “inside”, “outside”, “upper”, “lower”, etc. is based on the orientation or position relationship shown in the accompanying drawings. It is only for the convenience of describing the present invention and simplifying the description, and does not indicate or imply that the device or element referred to must have a specific orientation, be constructed and operated in a specific orientation. Therefore, it cannot be understood as a limitation of the present invention. When the absolute position of the described object changes, the relative position relationship may also change accordingly. In the present invention, unless otherwise clearly stipulated and limited, the terms such as “attachment” should be understood in a broad sense. For example, it can be a fixed connection, a detachable connection, or an integral whole; it can be directly connected or indirectly connected through an intermediate medium, and it can be the internal connection of two elements or the interaction relationship between two elements. For ordinary technicians in this field, the specific meanings of the above terms in the present invention can be understood according to the specific circumstances.

[0072] Example

[0073] In this example, data was collected from 53 real patients at Beth Israel Deaconess Medical Center. For each patient, 8 minutes of PPG signal and blood oxygen saturation index were collected, with a PPG signal sampling rate of 125 Hz and a blood oxygen saturation sampling rate of 1 Hz.

[0074] Before filtering the PPG signal, the collected PPG signal was segmented and intercepted using a sliding window of 10 seconds, with each sliding interval of 5 seconds, to obtain a total of 5035 samples;

[0075] In order to obtain reliable prediction results and avoid data leakage, the samples were divided into training and test sets according to patient pairs. 42 patients formed the training set and 11 patients formed the test set to evaluate the performance of the algorithm on unseen patients.

[0076] The present invention provides a method and system for estimating blood oxygen saturation based on a PPG signal, comprising the following steps:

[0077] S1: The PPG signal is an indirect measurement signal. Its accuracy is greatly affected by the measurement environment, measurement equipment, measurement method, and other components in human tissue. Its effective components are mainly concentrated in the frequency range of heart rate and above. The signal components below the heart rate are baseline drift. Therefore, a high-pass filter is required to filter out the signal components below the heart rate and smooth the PPG signal. In addition, the sensor or external factors may introduce high-frequency noise into the PPG signal, affecting feature extraction. Therefore, a low-pass filter is required to filter out high-frequency noise.

[0078] However, motion artifacts that fall within the effective frequency of the PPG signal cannot be filtered out by simple high-pass or low-pass filters. To address this issue, we can start by observing the frequency domain characteristics of the PPG signal and design a filter based on the characteristics of the PPG signal. Generally, various physiological indicators of the human body, such as systolic blood pressure, diastolic blood pressure, and blood oxygen saturation, are relatively stable in a short period of time. Therefore, within a cardiac cycle, the waveform of the PPG signal should be highly similar. In the frequency domain, this similarity is manifested in that the main components of the PPG signal are distributed on the multiples of the heart rate. The denoising effect of the comb filter on the multiples signal makes it very suitable for noise removal of PPG signals.

[0079] A bandpass filter is designed based on the Butterworth filter, and a comb filter is designed based on the characteristics of motion artifacts in the PGG signal. The motion artifacts and noise in the PGG signal are removed by the bandpass filter and the comb filter.

[0080] In step S1, a bandpass filter is designed based on a Butterworth filter, specifically including the following steps:

[0081] Step 1: Filter out the signals below the heart rate ω through a high-pass filter HR PPG signal components, the order of the high-pass filter is n h Set to 2, the cutoff frequency of the high-pass filter ω ch Set to 0.5ω HR , high-pass filter at heart rate ω HR The gain at the heart rate is 97.0%, and the high-pass filter is at the heart rate ω HR When the gain is greater than 95%, the order n h With the cutoff frequency ω ch Satisfy between:

[0082]

[0083] Among them H c (jω HR ) represents the high-pass filter at heart rate ω HR Gain at

[0084] Step 2: Filter out high-frequency noise through a low-pass filter. The order of the low-pass filter is n. l If it is set to 2, there is no strict limit on the upper frequency limit of the PPG signal. As the frequency increases, the amplitude of the PPG signal gradually decreases. When it reaches 10Hz, the amplitude is basically negligible. Therefore, the cutoff frequency of the low-pass filter ω ch When set to 10Hz, the frequency of the effective information of the PPG signal is lower than 8Hz, and the gain of the low-pass filter at 8Hz is 84.2%;

[0085] Step 3: Connect the high-pass filter and the low-pass filter in series to obtain a band-pass filter. The order of the band-pass filter is 2, the high-pass frequency is 0.5 times the heart rate, and the low-pass frequency is 10Hz. The transfer function of the band-pass filter is set to:

[0086]

[0087] Among them H bp (jω) represents the frequency response function of the bandpass filter, H hp (jω) represents the frequency response function of the high-pass filter, H lp (jω) represents the frequency response function of the low-pass filter.

[0088] In step S1, the base frequency of the comb filter is set to the heart rate ω HR , estimate the heart rate ω through frequency domain analysis algorithm HR , specifically including the following steps:

[0089] Step 1: Convert the original PPG signal X and the original sampling frequency and heart rate estimation range Input frequency domain analysis algorithm, heart rate estimation range Determined through experience;

[0090] Step 2: Perform FFT transformation on the original PPG signal X to obtain the heart rate estimation range The signal amplitude S at each frequency ω within x ;

[0091] Step 3: Calculate the average amplitude of the original PPG signal X at the frequency ω And the heart rate ω HR Set to the average amplitude of the octave The maximum value of the frequency ω, the average amplitude of the frequency doubling Set to:

[0092]

[0093] Where n represents an integer rounded down, a kω Indicates the gain at the kth frequency.

[0094] S2: Extract PGG signal features. The extraction results are shown in Table 1.

[0095] Table 1: PPG signal characteristics

[0096]

[0097] The PGG signal features were preliminarily screened by using the threshold filtering method based on the correlation coefficient;

[0098] Step S2 specifically includes the following steps:

[0099] S21: Calculate the Spearman rank correlation coefficient between the PPG signal and blood oxygen saturation;

[0100] S22: Given the high computational complexity of feature elimination and the need to train the model multiple times, other feature elimination methods should be used to preliminarily screen the features. In this embodiment, the Spearman rank correlation coefficient is used to preliminarily screen the PPG signal features and remove features below the threshold r. threshold The Spearman rank correlation coefficient corresponds to the PPG signal characteristics, the threshold r threshold Set to 0.1, Spearman rank correlation coefficient r s Set to:

[0101]

[0102] R(X) and R(Y) both represent random variable samples, σ R(X) and σ R(Y) Both represent the standard deviation of the random variable sample.

[0103] S3: Recursive feature elimination is a model-based feature selection method, which belongs to the parcel feature screening method. The main principle of recursive feature elimination is to recursively build a model and evaluate the importance of each feature, gradually eliminating the features that contribute the least to the model performance, thereby achieving the purpose of feature screening. The recursive feature elimination method is used to screen from the three dimensions of time domain, frequency domain and statistics to extract the PGG signal feature set X. selected ;

[0104] like Figure 1 As shown, step S3 specifically includes the following steps:

[0105] S31: Determine a base model, where the base model is set to be a linear regression model, a support vector machine model, or a random forest model. In this embodiment, the base model is set to be a random forest model with 100 trees.

[0106] S32: training the base model using PGG signal features;

[0107] S33: calculating the importance scores I of different PGG signal features using the base model, and ranking the importance scores I;

[0108] S34: Eliminate the PGG signal feature with the lowest importance score I, that is, the feature with the least impact on model performance;

[0109] S35: Retrain the base model using the PGG signal features after elimination. If the training result meets the performance evaluation index or feature quantity standard, extract the PGG signal feature set X corresponding to the training result. slected, otherwise repeat steps S31-S34.

[0110] like Figure 2 As shown, in step S35, in order to determine the number of features to be retained, the model should be trained and tested on each number of features, and the error of the model should be recorded. When the number of retained features is 1 to 20, the model is trained and the error is recorded respectively. As the number of features increases, the effective information obtained by the model gradually increases, so the error decreases rapidly. When the number of features is 6, the error reaches the lowest. Subsequently, the increase in invalid features causes the model to overfit, and the error gradually increases slowly. Therefore, when the number of features is set to 6, the performance of the model is the best;

[0111] The six features are the average systolic peak height, the duration of the diastolic peak at 33% of the height, the signal average, the signal upper quartile, the power P of 0-2Hz, and the average value of the diastolic peak at 33% of the height. [0-2Hz] , the ratio of the signal standard deviation to the median value r sd / m ;

[0112] 0-2Hz power P [0-2Hz] Set to:

[0113]

[0114] Where X(f) represents the Fourier transform of the signal;

[0115] The ratio of the signal standard deviation to the median value r sd / m Set to:

[0116]

[0117] where x i represents a single sample value, μ represents the mean value of the signal, and median represents the median value of the signal.

[0118] S4: The PGG signal feature set X is regressed by an integrated regressor based on the random forest algorithm. selected After estimating and obtaining the 6 best features, the random forest estimator is trained to estimate the blood oxygen saturation. During the estimation process, for each sample, the algorithm makes a decision on each decision tree. The random forest estimator consists of 1000 binary decision trees, each of which uses 5 features as input. The termination condition is that one sample remains for each node to better fit the data. The final estimation result is the average of the estimation results of all decision trees.

[0119] Step S4 specifically includes the following steps:

[0120] S41: Train N decision trees using N random subsets of the sample set;

[0121] S42: Recursively split the node, randomly select n f node features, and split the optimal features and optimal thresholds. The splitting stop condition is set to the node depth or the number of node samples. The feature with the smallest mean square error after splitting is selected as the optimal feature of the node.

[0122] S43: Record the optimal features and thresholds of all nodes to form a decision tree. N decision trees form a random forest.

[0123] S44: Starting from the root node of the decision tree, the decision is made based on the node's features and thresholds, and the node is transferred to the left or right child node. If the node is transferred to a leaf node, the estimated value of the leaf node is recorded.

[0124] S45: Calculate the average of the estimated values ​​of the leaf nodes of N decision trees to obtain the final estimated value Final estimate Set to:

[0125]

[0126] Where B represents the number of decision trees, and the number of decision trees B is set to 1000. b (x) represents the final estimated value of the bth decision tree;

[0127] S46: Combine performance evaluation indicators to evaluate the consistency between the estimated value and the true value, and use the support vector machine algorithm, decision tree algorithm and multi-layer perceptron algorithm for comparative testing.

[0128] In step S46, the performance evaluation indicators include mean absolute error (MAE) and mean square error (MSE). The calculation methods of mean absolute error (MAE) and mean square error (MSE) are:

[0129]

[0130] N s represents the total number of samples, y j Indicates the predicted value of blood oxygen, Indicates the actual value of blood oxygen.

[0131] In step S46, the consistency between the estimated value and the true value is evaluated by the Bland-Altman plot. The horizontal axis of the Bland-Altman plot is set to the average of the predicted value and the estimated value, and the vertical axis of the Bland-Altman plot is set to the difference between the estimated value and the true value. The basic idea is to use a scatter plot to show the difference between the two measurement methods of the same indicator. If most of the points in the plot fall within plus or minus 1.96 times the standard deviation, it means that the consistency of the two measurement methods is good.

[0132] The Brand-Altman diagram drawn from the test results is as follows Figures 3 to 7 shown.

[0133] like Figure 3 As shown in the figure, the standard deviation of the difference between the estimated value and the true value of the random forest algorithm is 1.71, and only 3.63% of the test data fall outside the range of plus or minus 1.96 times the standard deviation, indicating that the blood oxygen saturation measured by this method is well consistent with the true value of blood oxygen saturation. At the same time, the average value of the difference of the test data is -0.046, indicating that the estimated value of the test data has basically no overall offset;

[0134] Several algorithms used for comparative testing showed large errors:

[0135] like Figure 4 As shown, the support vector machine algorithm cannot fit the training data well. 8.95% of the test data exceeds the standard deviation of ±1.96 times. Although the average value of the difference is only 0.06, indicating that there is no obvious overall deviation in the estimated value, the standard deviation reaches 2.03.

[0136] like Figure 5 As shown in the figure, in order to prevent the decision tree from overfitting, the node is manually set to stop splitting when it reaches 15 layers, so that the training data can obtain a more appropriate fitting level. However, the difference between its estimated value and the true value is more dispersed, and the standard deviation of the difference value rises to 2.30, which is 35.3% higher than the random forest method.

[0137] For the multi-layer perceptron algorithm, in order to compare the effect differences between different numbers of nodes, two structures were used for testing respectively;

[0138] like Figure 6 As shown in the figure, the multilayer perceptron consisting of four layers of 100 neurons cannot fit the training data well after 200 iterations, and the difference values ​​of the test data also show a large offset and dispersion, with an average of -1.80 and a standard deviation of 2.14;

[0139] like Figure 7 As shown in the figure, in order to improve the fitting ability of the multilayer perceptron, the number of neurons in the 1st and 4th layers was increased to 200, and the number of neurons in the 2nd and 3rd layers was increased to 400. The training data was well fitted, and the difference between the estimated value and the true value of most training data dropped to within 2.5. However, the difference value of the test data became more dispersed, indicating overfitting. The standard deviation of the estimated value increased to 2.41, and the average value reached -0.84.

[0140] After the above tests, the performance evaluation indicators of different methods on the test data can be obtained as shown in Table 2;

[0141] Table 2: Performance evaluation metrics of different methods

[0142]

[0143] The random forest algorithm achieved the best performance in both indicators, which further demonstrated the effectiveness and accuracy of the random forest algorithm.

[0144] A blood oxygen saturation estimation system based on PPG signals, comprising a motion artifact removal module, a signal feature extraction module and a blood oxygen saturation estimation module;

[0145] The motion artifact removal module includes a noise removal algorithm based on a bandpass filter designed with a Butterworth filter, and a comb filter designed based on the characteristics of motion artifacts in PPG signals, which can remove motion artifacts from the signal;

[0146] The signal feature extraction module can extract a variety of PGG signal features from three dimensions: time domain, frequency domain, and statistics, which improves the information density of the input data and facilitates the subsequent processing of the estimation model;

[0147] The blood oxygen saturation estimation module contains an integrated regressor based on the random forest algorithm, which estimates the blood oxygen saturation of the human body based on multiple features obtained by the signal feature extraction module.

[0148] Therefore, the present invention adopts the above-mentioned blood oxygen saturation estimation method and system based on PPG signals to monitor the dynamic changes of blood oxygen saturation in real time, improve the generalization ability and prediction accuracy of the prediction model, and provide more accurate blood oxygen saturation estimation.

[0149] Finally, it should be noted that the above embodiments are only used to illustrate the method scheme of the present invention and not to limit it. Although the present invention has been described in detail with reference to the preferred embodiments, ordinary method personnel in this field should understand that they can still modify or replace the method scheme of the present invention with equivalents, and these modifications or equivalent replacements cannot cause the modified method scheme to deviate from the spirit and scope of the method scheme of the present invention.

Claims

1. A method for estimating blood oxygen saturation based on PPG signals, characterized in that: The following steps are involved: S1: Design a bandpass filter based on the Butterworth filter, and design a comb filter based on the characteristics of motion artifacts in the PPG signal. Use the bandpass filter and comb filter to remove motion artifacts and noise in the PPG signal. S2: Extract PPG signal features and perform preliminary screening of PPG signal features through threshold filtering method based on correlation coefficient; S3: Extract the PPG signal feature set by performing parcel feature screening from the time domain, frequency domain, and statistical dimensions through recursive feature elimination ; Step S3 specifically includes the following steps: S31: Determine a base model, which is set to a linear regression model, a support vector machine model, or a random forest model; S32: training a base model using PPG signal features; S33: Calculate the importance scores of different PPG signal features through the base model , score the importance Sort by S34: Eliminate importance score Minimum PPG signal characteristics; S35: Retrain the base model using the PPG signal features after elimination. If the training result meets the performance evaluation index or feature quantity standard, extract the PPG signal feature set corresponding to the training result. , otherwise repeat steps S31-S34; In step S35, the number of features is set to 6, and the 6 features are the average systolic peak height, the duration of the diastolic peak at 33% height, the signal average, the signal upper quartile, the power of 0-2Hz, and the average value of the systolic peak height. , the ratio of the signal standard deviation to the median ; 0-2Hz power Set to: in represents the Fourier transform of the signal, Indicates the number of signal samples; Ratio of the signal standard deviation to the median Set to: in represents a single sample value, represents the average value of the signal, represents the median value of the signal; S4: Based on the PPG signal feature set, an integrated regressor based on the random forest algorithm is used Estimate blood oxygen saturation.

2. The method for estimating blood oxygen saturation based on PPG signals according to claim 1, wherein: In step S1, a bandpass filter is designed based on a Butterworth filter, specifically including the following steps: Step 1: Filter out the pulses below the heart rate using a high-pass filter PPG signal components, the order of the high-pass filter Set to 2, the cutoff frequency of the high-pass filter Set to , high-pass filter in heart rate The gain at the heart rate is 97.0%, and the high-pass filter is When the gain is greater than 95%, the order and cutoff frequency Satisfy between: in Indicates that the high-pass filter is applied to the heart rate Gain at Step 2: Filter out high-frequency noise through a low-pass filter. The order of the low-pass filter is Set to 2, the cutoff frequency of the low-pass filter When set to 10 Hz, the frequency of the effective information of the PPG signal is lower than 8 Hz, and the gain of the low-pass filter at 8 Hz is 84.2%; Step 3: Connect the high-pass filter and the low-pass filter in series to obtain a band-pass filter. The transfer function of the band-pass filter is set to: in represents the frequency response function of the bandpass filter, represents the frequency response function of the high-pass filter, represents the frequency response function of the low-pass filter, Indicates the frequency of the input signal.

3. The method for estimating blood oxygen saturation based on PPG signals according to claim 1, wherein: In step S1, the base frequency of the comb filter is set to the heart rate , heart rate estimation through frequency domain analysis algorithm , specifically including the following steps: Step 1: Convert the original PPG signal X and the original sampling frequency and heart rate estimation range - Input frequency domain analysis algorithm, heart rate estimation range - Determined through experience; Step 2: Perform FFT transformation on the original PPG signal X to obtain the heart rate estimation range - Each frequency The signal amplitude on ; Step 3: Calculate the original PPG signal X at frequency The average amplitude of the frequency doubling and the heart rate Set to the average amplitude of the octave The frequency corresponding to the maximum value , the average amplitude of the double frequency Set to: in represents an integer rounded down, represents the floor operator, Indicates the Gain at the frequency doubling.

4. The method for estimating blood oxygen saturation based on PPG signals according to claim 1, wherein: Step S2 specifically includes the following steps: S21: Calculate the Spearman rank correlation coefficient between PPG signal characteristics and blood oxygen saturation; S22: Preliminary screening of PPG signal features using the Spearman rank correlation coefficient to remove signals below the threshold The Spearman rank correlation coefficient corresponding to the PPG signal characteristics, threshold Set to 0.1, Spearman rank correlation coefficient Set to: Represents and are random variable samples, and Both represent the standard deviation of the random variable sample.

5. The method for estimating blood oxygen saturation based on PPG signals according to claim 1, wherein: Step S4 specifically includes the following steps: S41: Through the sample set Random subset training decision trees; S42: Recursively split the nodes and randomly select node features, and split the optimal features and optimal thresholds. The splitting stop condition is set to the node depth or the number of node samples. The feature with the smallest mean square error after splitting is selected as the optimal feature of the node. S43: Record the optimal features and thresholds of all nodes to form a decision tree. A random forest is composed of decision trees; S44: Starting from the root node of the decision tree, the decision is made based on the node's features and thresholds, and the node is transferred to the left or right child node. If the node is transferred to a leaf node, the estimated value of the leaf node is recorded. S45: Calculation The average of the estimated values ​​of the leaf nodes of the decision tree is used to obtain the final estimated value , the final estimate Set to: in Indicates the number of decision trees, the number of decision trees Set to 1000, Indicates the The final estimated value of the decision tree; S46: Combine performance evaluation indicators to evaluate the consistency between the estimated value and the true value, and use the support vector machine algorithm, decision tree algorithm and multi-layer perceptron algorithm for comparative testing.

6. The method for estimating blood oxygen saturation based on PPG signals according to claim 5, wherein: In step S46, the performance evaluation indicators include mean absolute error (MAE) and mean square error (MSE). The calculation methods of mean absolute error (MAE) and mean square error (MSE) are: represents the total number of samples, Indicates the predicted value of blood oxygen, Indicates the actual value of blood oxygen.

7. The method for estimating blood oxygen saturation based on PPG signals according to claim 5, wherein: In step S46, the consistency between the estimated value and the true value is evaluated by the Bland-Altman plot, the horizontal axis of the Bland-Altman plot is set to the average of the predicted value and the estimated value, and the vertical axis of the Bland-Altman plot is set to the difference between the estimated value and the true value.

8. A system for estimating blood oxygen saturation based on a PPG signal according to any one of claims 1 to 7, characterized in that: It includes a motion artifact removal module, a signal feature extraction module and a blood oxygen saturation estimation module.

Images