Wrist PPG signal quality evaluation method based on LightGBM model

Abstract

The invention relates to a wrist PPG signal quality evaluation method based on a LightGBM model, and belongs to the field of biomedical signal analysis. The method comprises the following steps: performing signal segment segmentation on an acquired wrist PPG signal to obtain a wrist PPG signal sample, performing PPG signal preprocessing on the wrist PPG signal sample, extracting time-domain and frequency-domain mathematical statistical characteristics and similarity characteristics of the wrist PPG signal and a template PPG, performing signal quality category label distribution, constructing a data set, training a Light GBM model according to the data set, and performing PPG signal quality classification label distribution. And inputting the feature data of the wrist PPG signal subjected to feature extraction into the constructed LightGBM model so as to output the signal quality of the wrist PPG signal. According to the method, the signal quality of the wrist PPG signal collected under the complex wrist motion state and illumination condition can be evaluated, and the accuracy and precision of wrist PPG signal quality evaluation are effectively improved.

Description

Technical Field

[0001] The present invention belongs to the field of biomedical signal analysis, and relates to a method for evaluating the quality of wrist PPG signals based on the LightGBM model. Background Art

[0002] With the increasing awareness of public health, more and more wearable devices supporting vital sign detection have emerged in the market, such as smart bracelets, health watches, etc. These devices have extended the application of PPG technology from clinical applications to daily life.

[0003] This technology uses the absorption and scattering characteristics of light to detect blood dynamics. By emitting light to irradiate the skin and using a detector to capture the light reflected or transmitted through the blood, the light intensity signal is analyzed to reflect the blood flow characteristics. By analyzing these signals, non-invasive detection of vital sign parameters such as heart rate, blood oxygen saturation, blood pressure, and cardiac output can be achieved.

[0004] However, the wrist PPG signals detected by wearable devices are often mixed with serious interference signals due to the wearer's daily activities and complex lighting conditions in the environment, which affects the accuracy of further calculation of vital sign parameters and is prone to false alarms when monitoring vital sign parameters such as blood oxygen saturation. Therefore, it is a very important and meaningful step to determine the signal quality before analyzing the wrist PPG signals. Summary of the Invention

[0005] In view of this, the purpose of the present invention is to provide a method for evaluating the quality of wrist PPG signals based on the LightGBM model, which is applied to an electronic device equipped with a wrist PPG acquisition function to improve the accuracy of wrist PPG signals for analyzing vital sign parameters such as heart rate, blood oxygen saturation, blood pressure, and cardiac output.

[0006] To achieve the above purpose, the present invention provides the following technical solutions:

[0007] A method for evaluating the quality of wrist PPG signals based on the LightGBM model specifically includes the following steps:

[0008] S1: Use a wearable PPG detection electronic device to collect wrist PPG signals;

[0009] S2: Perform signal segmentation on the collected original wrist PPG signals with a specific window length to obtain target wrist PPG signal samples;

[0010] S3: Preprocess the target wrist PPG signal samples to obtain preprocessed wrist PPG signal segments;

[0011] S4: Assign signal quality category labels to the preprocessed wrist PPG signal segments;

[0012] S5: Extract features from the preprocessed wrist PPG signal segments;

[0013] S6: Combine the wrist PPG signal feature dataset and the quality category labels of the preprocessed wrist PPG signal segments to obtain a wrist PPG signal quality evaluation dataset;

[0014] S7: Train a LightGBM model to obtain a wrist PPG signal quality evaluation model;

[0015] S8: According to the wrist PPG signal quality evaluation LightGBM model, input the data after feature extraction processing into the optimal LightGBM model in a fixed format to determine the quality category of the wrist PPG signal to be evaluated.

[0016] Further, step S3 specifically includes the following steps:

[0017] S31: Perform wavelet decomposition, denoising, and reconstruction processing on the target wrist PPG signal samples;

[0018] S32: Perform normalization processing on the reconstructed signal;

[0019] S33: Divide the normalized signal into pulse wave beats.

[0020] Further, step S4 specifically includes: According to the waveform diagram of the preprocessed wrist PPG signal segments, assign the category labels to the waveforms of the preprocessed wrist PPG signals within the preset signal quality level classification standard range.

[0021] The preset signal quality level classification standard range includes high-quality wrist PPG signals and low-quality wrist PPG signals.

[0022] Further, step S5 specifically includes:

[0023] (1) Extract the time-domain and frequency-domain mathematical statistical features of the preprocessed wrist PPG signal segments, including at least one of the following: skewness, kurtosis, Shannon entropy of pulse wave beats, perfusion index, maximum power, frequency corresponding to the maximum power, time span of pulse wave beats, peak value of pulse wave beats, number of beats in the preprocessed wrist PPG signal segments, Shannon entropy of the preprocessed wrist PPG signal segments, number of zero-crossing points;

[0024] (2) Extract the similarity features between the preprocessed wrist PPG signal segments and the wrist pulse wave template signals, including at least one of the following: Euclidean distance, correlation coefficient.

[0025] Further, step S7 specifically includes the following steps:

[0026] S71: Divide the wrist PPG signal quality assessment data set into a training set and a test set according to a certain ratio;

[0027] S72: Use the cross-validation grid search algorithm to perform hyperparameter search to obtain the optimal hyperparameters;

[0028] S73: Use the test set data to iteratively train the LightGBM model;

[0029] S74: Evaluate the classification effect of the LightGBM model according to the test set data;

[0030] S75: Retain the optimal LightGBM model;

[0031] S76: If the number of iterations reaches the maximum or the model performance does not improve within a fixed number of iterations, end the training and save the optimal model, otherwise return to step S72.

[0032] The beneficial effects of the present invention are as follows: The present invention makes full use of the time-frequency domain mathematical statistical characteristics of the wrist pulse wave signal, the similarity characteristics between the PPG beat and the PPG template signal, and the LightGBM model to realize the quality assessment of the wrist PPG signal. It has a small amount of calculation, requires a small number of training set samples, improves the accuracy and precision of the wrist PPG signal quality assessment, and has the real-time performance and reliability for use in engineering.

[0033] Other advantages, objectives and features of the present invention will be described to some extent in the subsequent description, and to some extent, will be obvious to those skilled in the art based on the study of the following text, or can be taught from the practice of the present invention. The objectives and other advantages of the present invention can be realized and obtained through the following description. BRIEF DESCRIPTION OF THE DRAWINGS

[0034] In order to make the objectives, technical solutions and advantages of the present invention clearer, the present invention will be described in detail preferably with reference to the accompanying drawings, where:

[0035] Figure 1 is the implementation flowchart of the wrist PPG signal quality assessment method based on the LightGBM model of the present invention;

[0036] Figure 2 is the training flowchart of the LightGBM model. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0037] The following describes the embodiments of the present invention by specific examples, and those skilled in the art can easily understand other advantages and effects of the present invention from the contents disclosed in this specification. The present invention can also be implemented or applied through other different specific embodiments, and the details in this specification can also be modified or changed in various ways based on different viewpoints and applications without departing from the spirit of the present invention. It should be noted that the illustrations provided in the following embodiments only illustrate the basic concept of the present invention in a schematic manner, and the following embodiments and features in the embodiments can be combined with each other without conflict.

[0038] See also Figures 1 to 2 The present invention proposes a wrist PPG signal quality assessment method based on the LightGBM model, which can identify the quality of wrist PPG signals collected by a wearable PPG detection device under complex wrist motion states and lighting conditions, effectively improve the accuracy and precision of wrist PPG signal quality assessment, thereby improving the accuracy of vital sign parameter analysis using the PPG signal. The preferred example process is as follows Figure 1 As shown, the process specifically includes:

[0039] Step 1: Collect wrist PPG signals using electronic equipment.

[0040] Step 2: Perform signal segmentation of a specific window length on the collected original wrist PPG signal to obtain target wrist PPG signal samples.

[0041] Step 3: Preprocess the target wrist PPG signal sample to obtain the preprocessed wrist PPG signal fragment, which specifically includes the following steps:

[0042] Step 31: Optimize the "sym4" wavelet basis, perform 9-layer decomposition on the target wrist PPG signal sample, adopt hard threshold denoising with a fixed threshold, and then perform wavelet reconstruction to remove high-frequency noise and baseline drift;

[0043] Step 32: normalize the target wrist PPG signal samples to eliminate dimension effects;

[0044] Step 33: Divide the target wrist PPG signal samples into beats.

[0045] Step 4: Assign signal quality category labels to the preprocessed wrist PPG signal segments.

[0046] Step 5: Perform feature extraction on the preprocessed wrist PPG signal segment, specifically including:

[0047] (1) Extracting mathematical statistical features of wrist PPG signal segments in the time and frequency domain after processing, including but not limited to:

[0048] Skewness, kurtosis, Shannon entropy of pulse wave beats, perfusion index, maximum power, frequency corresponding to the maximum power, time span of pulse wave beats, peak value of pulse wave beats, number of beats in the preprocessed wrist PPG signal segment, Shannon entropy of the preprocessed wrist PPG signal segment, number of zero-crossings.

[0049] Skewness: Skewness, as a statistic for measuring the asymmetry of signal distribution, describes the skewness direction and degree of signal distribution. For signal X(t), skewness is defined as:

[0050]

[0051] where E[·] is the expected value, μ is the mean, and σ is the standard deviation.

[0052] Kurtosis: Kurtosis is a statistic for measuring the "tail" degree of signal distribution, also known as peakness. It describes the peak distribution of signal distribution. For signal X(t), kurtosis is defined as:

[0053]

[0054] Shannon Entropy: Shannon Entropy, as an index for measuring signal uncertainty or information content, is defined as:

[0055] Shannon Entropy=-∑p(x i )logp(x i )

[0056] where p(x i ) is the probability of obtaining the value of x i .

[0057] Perfusion Index: The perfusion index (PI) is defined as the ratio of the pulsatile component (AC, mainly pulsatile arterioles) to the non-pulsatile component (DC, including veins, capillaries, non-pulsatile arterial blood, and tissue) of local tissue at a specific wavelength, and is defined as:

[0058]

[0059] where PS is the pulsating signal and NS is the non-pulsating signal.

[0060] Maximum Power Pmax and Its Corresponding Frequency f: The maximum power refers to the maximum amplitude value in the signal power spectrum, and the corresponding frequency refers to the frequency point where this maximum power value is located. The spectrum of the signal can be obtained through Fourier transform, and then the point with the maximum amplitude in the spectrum and its corresponding frequency value can be found.

[0061] Beat Interval Time: The beat interval time refers to the time span from the starting point to the ending point of the pulse wave, that is:

[0062] ΔT = t end -t start

[0063] Peak-to-peak value of the beat: The peak-to-peak value of the beat refers to the amplitude difference between the peak point and the starting point of the pulse wave beat after normalization.

[0064] Number of beats: The number of beats refers to the number of complete beats detected after beat division in the 3s wrist photoplethysmogram data.

[0065] Overall data entropy value: In this method, this index is defined as the Shannon entropy of the overall signal segment data.

[0066] Number of zero crossings: It refers to the number of times the signal value changes from positive to negative or from negative to positive within a certain time interval, that is, the number of times the signal curve crosses the zero horizontal line. In a continuous-time signal, this usually means the points where the signal curve intersects the zero value line. In a discrete-time signal (such as a digital signal), zero crossings occur when the sign changes between adjacent samples.

[0067] (2) Extract the similarity features between the preprocessed wrist PPG signal segment and the wrist pulse wave template signal, including but not limited to: Euclidean distance, correlation coefficient.

[0068] The formula for calculating the Euclidean distance is as follows,

[0069] d(x i , y i ) = |x i - y i |

[0070] The formula for calculating the correlation coefficient is as follows:

[0071]

[0072] In the formula, r XY is the correlation coefficient between sequences X and Y, x i and y i are the i-th sample values of sequences X and Y respectively, and are the sample means of sequences X and Y respectively, and n is the number of samples. When r XY = 1, it means that the two sequences are completely positively correlated; when r XY = -1, it means that the two sequences are completely negatively correlated; when r XY = 0, it means that the two sequences have no linear relationship.

[0073] Step 6: Combine the quality labels of the preprocessed wrist PPG signal segments in the wrist PPG signal feature dataset to obtain the wrist PPG signal quality evaluation dataset.

[0074] Step 7: Train the LightGBM model to obtain a wrist PPG signal quality assessment model, specifically including the following steps:

[0075] Step 71: Divide the wrist PPG signal quality assessment dataset into a training set and a test set according to a certain ratio;

[0076] Step 72: Perform hyperparameter search through the cross-validation grid search algorithm to obtain the optimal hyperparameters;

[0077] Step 73: Iteratively train the LightGBM model using the test set data;

[0078] Step 74: Evaluate the classification effect of the LightGBM model based on the test set data;

[0079] Step 75: Considering multiple aspects such as robustness and stability, retain the optimal LightGBM model in multiple iterations;

[0080] Step 76: If the number of iterations reaches the set maximum number of iterations or the model performance does not improve within a fixed number of iterations, end the training and save the best model. Otherwise, return to Step 72.

[0081] Step 8: According to the wrist PPG signal quality assessment LightGBM model, input the data after feature extraction processing into the optimal LightGBM model in a fixed format to determine the quality category of the wrist PPG signal to be evaluated.

[0082] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit them. Although the present invention has been described in detail with reference to the preferred embodiments, those of ordinary skill in the art should understand that the technical solutions of the present invention can be modified or equivalently replaced without departing from the spirit and scope of the present technical solution, and they should all be covered within the scope of the claims of the present invention.

Claims

1. A wrist PPG signal quality assessment method based on the LightGBM model, characterized in that: The method specifically comprises the following steps: S1: Use wearable PPG detection electronic equipment to collect wrist PPG signals; S2: performing signal segmentation of a specific window length on the collected original wrist PPG signal to obtain a target wrist PPG signal sample; S3: preprocessing the target wrist PPG signal sample to obtain a preprocessed wrist PPG signal segment; S4: assigning signal quality category labels to the preprocessed wrist PPG signal segments; S5: extracting features from the preprocessed wrist PPG signal segments; S6: combining the wrist PPG signal feature dataset and the preprocessed wrist PPG signal segment quality category label to obtain a wrist PPG signal quality assessment dataset; S7: Train the LightGBM model to obtain the wrist PPG signal quality assessment model; S8: Evaluate the LightGBM model according to the quality of the wrist PPG signal, input the data after feature extraction processing into the optimal LightGBM model in a fixed format, and determine the quality category of the wrist PPG signal to be evaluated.

2. The wrist PPG signal quality assessment method according to claim 1, characterized in that: Step S3 specifically includes the following steps: S31: performing wavelet decomposition, denoising and reconstruction processing on the target wrist PPG signal sample; S32: performing standardization processing on the reconstructed signal; S33: dividing the normalized signal into pulse wave beats.

3. The wrist PPG signal quality assessment method according to claim 1, characterized in that: Step S4 specifically includes: according to the preprocessed wrist PPG signal segment waveform diagram, assigning a category label to the preprocessed wrist PPG signal waveform within a preset signal quality level classification standard range.

4. The wrist PPG signal quality assessment method according to claim 1, characterized in that: Step S5 specifically includes: (1) extracting mathematical statistical features of the wrist PPG signal segment in the time domain and frequency domain after preprocessing, including at least one of the following: skewness, kurtosis, pulse wave beat Shannon entropy, perfusion index, maximum power, maximum power corresponding frequency, pulse wave beat time span, pulse wave beat peak, the number of beats in the wrist PPG signal segment after preprocessing, Shannon entropy of the wrist PPG signal segment after preprocessing, and the number of zero crossings; (2) extracting similarity features between the preprocessed wrist PPG signal segment and the wrist pulse wave template signal, including at least one of the following: Euclidean distance and correlation coefficient.

5. The wrist PPG signal quality assessment method according to claim 1, characterized in that: Step S7 specifically includes the following steps: S71: Divide the wrist PPG signal quality assessment dataset into a training set and a test set in proportion; S72: Use cross-validation grid search algorithm to perform hyperparameter search to obtain the optimal hyperparameter; S73: Iteratively train the LightGBM model using the test set data; S74: Evaluate the classification effect of the LightGBM model based on the test set data; S75: retain the optimal LightGBM model; S76: If the number of iterations reaches the maximum or the model performance does not improve within a fixed number of iterations, end the training and save the optimal model. Otherwise, return to step S72.

Images