Sleep state snoring prediction method based on pulse wave physiological indexes in waking state

Abstract

The invention discloses a sleep state snoring prediction method based on pulse wave physiological indexes in a waking state, which is used for predicting whether a subject snoring at night by recording pulse wave data of different breathing behaviors in the waking state and extracting the physiological indexes from the pulse wave data so as to solve the problem that current snoring evaluation needs to be performed in the sleep state. The method comprises the following steps: recording pulse wave signals during natural breathing, mouth breathing and single-nose breathing in a waking state; carrying out denoising and periodic validity evaluation on the signal; feature extraction is carried out based on the effective pulse wave period; the method comprises the following steps: by taking pulse wave characteristics of a sober subject as input characteristics, inquiring whether the subject snore or not during sleeping as a sample label, and constructing a training set; the classifier is trained; for the trained classifier, test data of a new user in a waking state are input, and whether the new user snore or not during sleeping can be predicted. According to the invention, a prediction method under a day waking state can be provided for evaluating whether snoring occurs during sleep or not.

Description

Technical Field

[0001] The present invention belongs to the technical field of information processing, and particularly relates to a method for predicting snoring during sleep based on physiological indicators of pulse wave in a waking state. Background Art

[0002] Snoring is a common sleep problem, which is a symptom manifested by poor breathing during sleep. Currently, snoring detection is mainly completed in two ways: an intuitive observation method of recording during night sleep, and a method of detecting sleepers using polysomnography during night sleep. The method of sleep recording has high requirements for the sound collection device and the device placement position; it requires people to be still and face the microphone during sleep, which is difficult to achieve in a non-laboratory environment; at the same time, there are various noise interferences, resulting in misjudgment of snoring recognition. The snoring detection method based on polysomnography requires expensive and large equipment for detection, and also requires the guidance of professional doctors. At the same time, both of the above two methods need to be carried out during the sleep state. Since sleep is a private behavior, it is difficult to be accepted by the public. Summary of the Invention

[0003] The main purpose of the present invention is to overcome the disadvantages and deficiencies of the existing snoring detection technology, and provide a method for predicting snoring during sleep based on physiological indicators of pulse wave in a waking state. This method has the characteristics of simple and inexpensive detection equipment, intuitive data sampling process, and no need for detection during night sleep. At the same time, it has a high prediction accuracy rate, and is an objective, simple method that can better protect user privacy, does not require night detection, and is more easily accepted by the public.

[0004] In order to achieve the above purpose, the present invention adopts the following technical solutions:

[0005] A method for predicting snoring during sleep based on physiological indicators of pulse wave in a waking state, aiming to improve the problems of expensive and complex equipment and the difficulty of acceptance in the traditional snoring measurement method that requires detection during sleep. The method for predicting snoring during sleep includes the following steps:

[0006] S1. Construct a test scenario. The test equipment consists of a pulse wave signal measurement sensor and a supporting upper computer. Find non-solitary subjects for data set collection. The subject presses the index finger on the sensor. The entire test process includes five steps: sitting still, natural breathing, mouth breathing, left nasal cavity breathing, and right nasal cavity breathing; record the pulse wave signals during the processes of natural breathing, mouth breathing, left nasal cavity breathing, and right nasal cavity breathing; at the same time, require the subject's roommate to listen during the subject's sleep that night to know whether the subject snores that night, and ask the subject whether he snores at night the next day. Use it as the training set sample label Label for supervised learning.

[0007] S2. Denoise each segment of the pulse wave signal collected from S1. Use low-pass filtering to remove high-frequency noise to obtain a smoothed pulse wave signal. Then, use the cubic spline interpolation baseline removal method to obtain the pulse wave signal after baseline removal. Next, perform a periodic validity assessment to remove signal cycles with poor signal quality, resulting in a high-quality pulse wave signal.

[0008] S3. For the high-quality pulse wave signal, extract the characteristic points of the pulse wave cycle, including the steepest rising point, starting point, and peak point. Based on the periodic characteristic points, extract the physiological characteristic data of the subject in this experiment, including heart rate (HR), heart rate variability (HRV), blood oxygen saturation (SpO2), and morphological index of the pulse wave (MI). Based on the smoothed pulse wave signal, perform frequency domain analysis to obtain the respiratory rate (RR). Using the pulse wave characteristics under natural breathing as the standard value, subtract the pulse wave characteristics during mouth breathing and single nasal cavity breathing respectively to obtain the physiological characteristic sequence.

[0009] S4. Use the physiological characteristic sequence of the subject when awake as the input feature, and whether the subject snores during sleep at night as the sample label to construct a training set.

[0010] S5. Use the training set obtained in step S4 to train the classifier.

[0011] S6. For new users who do not know whether they snore or not, when they are awake, repeat steps S1 - S3 to obtain user data. Different from the subjects, at this time, it is no longer required that the user knows whether they snore at night. Input the user data into the classifier to predict whether they snore during sleep.

[0012] Furthermore, the process of step S1 is as follows:

[0013] S101. Construct a test scenario. Under the conditions of suitable temperature and ventilation, find non-solitary subjects for data set collection. The subject presses their finger on the pulse wave signal sensor and sits back against the chair. The requirements for temperature and ventilation environment in this step are to ensure that the temperature and air environment will not interfere with the subject's experiment. The non-solitary condition ensures that when the subject sleeps at night, there will be a co-resident to observe whether they snore.

[0014] S102. Let the subject sit still for a duration of T 1 , so that the subject becomes calm and familiar with the test environment, avoiding interference with physiological signals due to the test environment. This step ensures that physiological signals will not be interfered with due to the subject's discomfort with the unfamiliar environment.

[0015] S103. Intervals T 2 Enable the fingers to fully relax. Since the subject's fingers need to be pressed on the pulse wave signal sensor during data collection, long-term compression of the fingers will affect blood flow and cause discomfort to the subject's fingers. Therefore, design the relaxation time T 2 After the subject's fingers are fully relaxed, conduct subsequent experimental procedures. After the fingers are fully relaxed, the subject breathes naturally, and pulse wave data is collected through the pulse wave signal sensor as the reference signal s raw (0), lasting for T 1 .

[0016] S104. Intervals T 2 After the fingers are fully relaxed, the subject breathes only with the oral cavity, collects pulse wave data, and obtains the pulse wave signal s raw (1) to evaluate the oral ventilation condition, lasting for T 1 , and the pulse wave signal s raw (1) obtained in this step will be used to extract pulse wave physiological indicators to obtain evaluation indicators for the ability of oral breathing.

[0017] S105. Intervals T 2 After the fingers are fully relaxed, the subject breathes only with the left nasal cavity, collects pulse wave data, and obtains the pulse wave signal s raw (2) to evaluate the ventilation condition of a single nasal cavity, lasting for T 1 , and the pulse wave signal s raw (2) obtained in this step will be used to extract pulse wave physiological indicators to obtain evaluation indicators for the ability of left nasal cavity breathing.

[0018] S106. Intervals T 2 After the fingers are fully relaxed, the subject breathes only with the right nasal cavity, collects pulse wave data, and obtains the pulse wave signal s raw (3) to evaluate the ventilation condition of a single nasal cavity, lasting for T 1 , and the pulse wave signal s raw (3) obtained in this step will be used to extract pulse wave physiological indicators to obtain evaluation indicators for the ability of right nasal cavity breathing.

[0019] The pulse wave signals obtained from the four processes are denoted as s raw (n), where n = 0, 1, 2, 3, corresponding to the four states of natural breathing, breathing only with the oral cavity, breathing only with the left nasal cavity, and breathing only with the right nasal cavity respectively.

[0020] S107. Require the subject's roommate to monitor during the subject's sleep on that night, and ask the subject whether he snores the next day. This result is obtained through actual observation by the roommate during the subject's sleep and is used as the training set sample label Label of the classifier, where Label ∈ {0, 1}.

[0021] Further, the process of step S2 is as follows:

[0022] S201. According to the effective frequency range of the pulse wave, perform low-pass filtering on each segment of the pulse wave signal. The effective range of the pulse wave signal is 0 - 10 Hz. Therefore, use a 10 Hz low-pass filter to perform signal filtering to remove high-frequency noise and obtain a smooth pulse wave signal s smooth (n), n = 0, 1, 2, 3. This step performs low-pass filtering and noise reduction on the pulse wave signal, providing a smooth pulse wave signal for the subsequent extraction of the cycle starting point of the pulse wave signal.

[0023] S202. Extract the starting point of each cycle of the signal. Use the local maximum extraction method to extract the starting point of each cycle. Based on the starting point, truncate the front and back of each segment of the signal to avoid incomplete cycles during sampling.

[0024] The starting point is obtained by calculating the local maximum point of the smooth pulse wave signal. For the starting point O i of the i-th cycle, there is s smooth [O i >s smooth [O i - d], where M is the number of sampling points within 1 s. Due to the characteristics of signal sampling, there may be multiple points with equal values at the local maximum. Therefore, when there are multiple adjacent local maximum points in the process of calculating the starting point, the local maximum point in the middle will be temporarily incremented by one to ensure that there are no multiple equal local maximum points during the calculation. The calculated local maximum point is the starting point O i of each cycle. Remove the first and last cycles of the pulse wave signal, as these cycles are often incomplete. After removal, the remaining number of cycles is K. In this step, the local maximum point in the middle is temporarily incremented by one to avoid interference from multiple equal local maximum points on the extraction of the starting point. The obtained cycle starting points will be used for cubic spline interpolation fitting calculation to remove the baseline of the pulse wave signal.

[0025] S203. Further, taking the starting point of each cycle of the smooth pulse wave as the base point, use the cubic spline interpolation method to fit the signal, subtract the signal, and obtain the pulse wave signal after baseline removal. For the pulse wave signal with K cycles, perform cubic spline interpolation function fitting.

[0026] The theory of cubic spline interpolation is as follows: For a general signal y[n], given the base points (n 0 , y 0 ), (n 1 , y 1 ), …, (n N , y N ), the length of each segmentation interval is:

[0027] h p = x p+1 - x p , p = 0, 1, …, N - 1

[0028] The fitting function for each segment is

[0029]

[0030] For p = 1, 2, …, N - 1, there are constraint conditions:

[0031]

[0032] And there are natural boundary conditions: c 0 = 0, c N = 0

[0033] After obtaining c through solution, coefficient calculation is performed

[0034] a p = y p

[0035]

[0036] According to the above algorithm, the starting points (O i , s smooth [O i ) of each period, and the end point (O K , s smooth [O K ) of the last period are used as the base points for cubic spline interpolation for fitting calculation, and the cubic spline fitting function can be obtained On this basis, signal baseline removal is performed, and there is:

[0037]

[0038] In actual operation, the end point of the last period will also be added to the cubic spline interpolation fitting for baseline removal of the last period. This step removes the baseline of the pulse wave signal, so that the subsequent evaluation of the cycle validity and feature extraction of the pulse wave signal will not be interfered by the baseline data.

[0039] S204. Perform cycle validity evaluation on each signal cycle of the signal after baseline removal, including skewness calculation and Pearson coefficient calculation, and remove the signal cycles where the Pearson coefficient and skewness do not meet the requirements to obtain a high-quality pulse wave signal s qualified .

[0040] Specifically, based on the starting point O i , i = 0, 1,..., K - 1, of the i-th cycle of the pulse wave signal, the signal s smooth is segmented by cycle to obtain K segments of single-cycle pulse wave signals, where the i-th cycle pulse wave signal is denoted as S i , i = 0, 1,..., K - 1.

[0041] Perform skewness calculation on the cycle segment S i . Specifically, perform normalization of the cycle sum on S i to obtain Y i . At this time, Y i is a sequence with a cycle sum of 1, and it is regarded as a probability distribution column for skewness calculation:

[0042]

[0043] where μ and σ are the expectation and variance of Y i respectively. Set the skewness threshold to θ s , where θ s is the average skewness of all cycle pulse waves. The pulse wave signal can be regarded as a right-skewed signal. Set when , the cycle segment S i is retained.

[0044] For each segment of single-cycle pulse wave signal after skewness screening, at this time, the proportion of effective signals in the pulse wave signal is relatively high, but there are still some invalid signals, and the Pearson coefficient calculation for adjacent cycles needs to be performed. Since the length of each cycle of the pulse wave signal is different, the number of sampling points for each cycle after cutting is inconsistent, and the Pearson correlation coefficient for adjacent cycles cannot be directly calculated. Therefore, cubic spline interpolation needs to be performed on the adjacent cycles obtained by cutting to obtain signals X i and X i+1 with the number of points being M, and then perform Pearson coefficient calculation:

[0045]

[0046] Set the Pearson coefficient threshold to θ p . When P i > θ p , the cycle segment S i is retained.

[0047] This step is used to extract the high-quality pulse wave signal period from the pulse wave signal after baseline removal, providing a high-quality pulse wave signal for subsequent feature points and feature extraction of the pulse wave, and ensuring the effectiveness of the extracted feature points and features.

[0048] Furthermore, the process of step S3 is as follows:

[0049] S301. For the high-quality pulse wave signal after removing the cycles with poor signal quality, extract the feature points of the pulse wave period. The feature points include the steepest rising point, starting point, and peak point. The steepest rising point is obtained by calculating the maximum point of the differential signal within each cycle, and the peak point is obtained by searching backward from the steepest rising point to find the first differential zero point. The feature points of the pulse wave period calculated in this step will be used for subsequent calculations of pulse wave morphology features, heart rate variability, and heart rate.

[0050] S302. Based on the feature points of the pulse wave period, calculate the heart rate, heart rate variability, and pulse wave morphology features; based on the smoothed pulse wave signal, use frequency domain analysis, and select the frequency value where the peak point in the frequency domain is located as the respiratory rate; the heart rate, heart rate variability, pulse wave morphology features, and respiratory rate together constitute the pulse wave feature sequence. Finally, ppg_feature(n) = {HR(n), HRV(n), SpO2(n), MI(n), RR(n)}, n = 0, 1, 2, 3, corresponding to four states of natural breathing, breathing only through the mouth, breathing only through the left nasal cavity, and breathing only through the right nasal cavity. The pulse wave feature sequences in the four states calculated in this step will be used to construct the left nasal cavity feature difference sequence, right nasal cavity feature difference sequence, and oral cavity feature difference sequence with the natural breathing pulse wave feature sequence as the standard value.

[0051] S303. For the pulse wave feature sequences ppg_feature(n) in the four stages obtained, n = 0, 1, 2, 3, with the pulse wave feature sequence ppg_feature(0) under natural breathing as the standard value, use the left nasal cavity pulse wave feature sequence to subtract the natural breathing pulse wave feature sequence to obtain the left nasal cavity feature difference sequence Bias L , use the right nasal cavity pulse wave feature sequence to subtract the natural breathing pulse wave feature sequence to obtain the right nasal cavity feature difference sequence Bias R ;

[0052] Bias L = ppg_feature(2) - ppg_feature(0)

[0053] Bias R = ppg_feature(3) - ppg_feature(0)

[0054] Left nasal cavity feature difference sequence Bias L , right nasal cavity feature difference sequence Bias R Take the maximum value of the absolute value of each feature difference in, and form a single-nose feature difference sequence Bias N . This construction method can obtain the maximum difference value between nasal breathing and natural breathing through this feature sequence when the obstruction degrees of the left and right nasal cavities are different. Specifically, the e-th element of sequence Bias N is:

[0055]

[0056] Use the pulse wave feature sequence under oral breathing and the natural breathing pulse wave feature sequence to perform a subtraction operation as the oral feature difference sequence Bias M .

[0057] Bias M = ppg_feature(1) - ppg_feature(0)

[0058] In this step, the left nasal cavity feature difference sequence Bias L , the right nasal cavity feature difference sequence Bias R , and the oral feature difference sequence Bias M are all obtained by subtracting the corresponding feature sequence from the natural breathing pulse wave feature sequence. The purpose is to remove the individual differences in the pulse wave features caused by other physiological factors of the subject that are not related to snoring. Concatenate the single-nose feature difference sequence Bias N and the oral feature difference sequence Bias M to obtain the sample input features for the training set, and use whether the known subject snores as the sample label to construct the data set.

[0059] Furthermore, the process of step S4 is as follows:

[0060] S401. For snoring subjects, their sample composition is: input feature Feature = {Bias N , Bias M}, sample label Label = {1}. The input features and sample labels obtained in this step together constitute the snoring subject samples, which will be used for the subsequent construction and training of the classifier.

[0061] S402. For non-snoring subjects, their sample composition is: input feature Feature = {Bias N , Bias M}, sample label Label = {0}. The input features and sample labels obtained in this step together constitute the non-snoring subject samples, which will be used for the subsequent construction and training of the classifier..

[0062] Further, the process of step S5 is as follows:

[0063] Using the training set data obtained in step S4, verify the classification results through K-fold cross-validation, and construct a snoring prediction classifier. The purpose of this step is to reduce the risk of model overfitting.

[0064] S501. For the data set obtained in step S4, perform feature selection during the actual training process. By selecting features based on the importance of random forests, eliminate features with less correlation to form an optimal subset of feature data.

[0065] S502. Split the optimal subset of feature data into a training set and a test set. Use the training set to train the classifier and the test set to test the performance of the classifier. Through this step of training the classifier, a classifier that can be used to predict snoring is obtained, and at the same time, a single performance evaluation of the classifier is performed.

[0066] S503. Use the method of K-fold cross-validation to repeat step S502 to obtain the average performance evaluation result of the classifier. In this step, K-fold cross-validation is a commonly used method for validating classification effects in classification problems. By repeatedly dividing the data set for retraining and validation, a comprehensive performance evaluation of the classifier is performed to avoid overestimating or underestimating the classification effect of the classifier due to unreasonable data set division. The classifier after effective average performance evaluation will be applied to the actual scenarios for users.

[0067] Further, the process of step S6 is as follows:

[0068] S601. For new users who do not know whether they snore at night, repeat steps S1 to S4, but different from the subjects during data set collection. In this step, the purpose of data collection is to predict whether the user snores or not, so at this time, it is no longer required that the user knows whether they snore at night; obtain the input features Feature={Bias N , Bias M} of the user. The data collection performed in this step will be used as the input of the already trained classifier to enable the classifier to predict whether the user snores at night.

[0069] S602. Input the input features of the new user into the classifier trained in step S5 to obtain the prediction result of whether the user snores. Through the input features of the user when awake, this step predicts whether the user snores at night, enabling the user to know whether they may snore at night while awake.

[0070] Compared with the prior art, the present invention has the following advantages and beneficial effects:

[0071] (1) The present invention provides a method for predicting snoring during sleep based on physiological indicators of pulse wave in the waking state. This method detects the characteristics of pulse wave signals under 4 processes of natural breathing, mouth breathing, left nasal cavity breathing, and right nasal cavity breathing of the user during the day through a pulse wave sensor, so as to predict whether snoring occurs at night. Among them, the method of detecting the fingertips through a pulse wave signal sensor has the advantages of simple and inexpensive detection equipment and intuitive data sampling process. The detection process based on the waking state during the day has the characteristic of not requiring detection during night sleep, and at the same time has a high prediction accuracy. It is an objective, simple method that can better protect the privacy of users, does not require night detection, and is more easily accepted by the public.

[0072] (2) The present invention proposes a method for serial screening of skewness - adjacent Pearson coefficient to evaluate the quality of pulse wave signal cycles. Using skewness, which is a necessary but not sufficient condition for high - quality signals, as a pre - screening link for high - quality signal cycles, the proportion of invalid cycles in the pulse wave signals to be detected is pre - reduced. The pre - screening effectively reduces the false alarm rate of subsequent cycle screening based on the Pearson coefficient of adjacent cycles. After the pre - screening of skewness is completed, the Pearson coefficient of adjacent cycles is further used to perform a cycle re - screening on the signal, effectively reducing the problem of too high a miss alarm rate existing in the skewness screening. Combining the skewness of signal cycles and the Pearson coefficient of adjacent cycles, this method has the advantages of lower miss alarm rate and false alarm rate compared with a single detection method.

[0073] (3) The present invention proposes an extraction of physiological characteristics of different breathing methods to eliminate individual differences. Independent experiments of natural breathing, left nasal cavity breathing, right nasal cavity breathing, and oral breathing of the subjects are designed and pulse wave data are collected. The pulse wave feature sequences ppg_feature(n), n = 0, 1, 2, 3 are extracted, corresponding to 4 states of natural breathing, breathing only with the mouth, breathing only with the left nasal cavity, and breathing only with the right nasal cavity respectively. On this basis, the difference is made with the pulse wave feature sequence of the subject's natural breathing as the benchmark to eliminate individual differences; in the processing of the feature differences between the left and right nasal cavities, the data of the left and right nasal cavities are screened based on the benchmark of "the largest absolute value" to avoid the influence of the data characteristics of the left and right nasal cavities due to symmetry on the classifier effect of snoring prediction; finally, the single - nose feature difference sequence Bias N , and the oral feature difference sequence Bias M are spliced to obtain the sample input feature Feature = {Bias N , Bias M} as the training set. This input feature is only caused by different breathing methods, effectively avoiding interference with the features caused by other physiological reasons of the subjects, thus affecting the prediction of snoring by the classifier.

[0074] (4) The present invention combines pulse wave characteristics, including heart rate HR, heart rate variability HRV, blood oxygen saturation SpO2, and pulse wave morphology index MI, and respiratory rate RR, to comprehensively evaluate whether a subject snores at night. During the actual construction of the classifier, for the sample input features Feature = {Bias N , Bias M}, through feature screening based on the importance of random forest, features with strong correlation with snoring are selected to form an optimal feature data subset for the training of the classifier, and are used to input the snoring prediction classifier to predict snoring. Compared with the detection scheme of a single index, it effectively avoids the one-sidedness of the results of single-index detection. BRIEF DESCRIPTION OF THE DRAWINGS

[0075] In order to more clearly illustrate the technical solutions in the embodiments of the present application, the following will briefly introduce the drawings required for the description of the embodiments. Obviously, the following drawings are only some embodiments of the present application. For those of ordinary skill in the art, without creative efforts, other drawings can also be obtained based on these drawings.

[0076] Figure 1 is a flowchart of a method for predicting snoring during sleep state based on physiological indicators of pulse wave in a waking state disclosed in the present invention;

[0077] Figure 2 is a flowchart of the preprocessing of the pulse wave signal in Embodiment 1 of the present invention;

[0078] Figure 3 is a flowchart of obtaining high-quality pulse waves by cycle screening in Embodiment 1 of the present invention;

[0079] Figure 4 is a flowchart of the process of extracting sample input features in Embodiment 1 of the present invention;

[0080] Figure 5 is a schematic diagram of the importance of the original data set to the random forest model in Embodiment 1 of the present invention;

[0081] Figure 6 is a schematic diagram of the importance of the sample features in the optimal feature data subset to the random forest model in Embodiment 1 of the present invention;

[0082] Figure 7 is a schematic diagram of the pulse wave signal before cycle screening in Embodiment 2 of the present invention;

[0083] Figure 8 is a schematic diagram of the pulse wave signal after cycle screening in Embodiment 2 of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0084] To enable those skilled in the art to better understand the solution of this application, the technical solution in the embodiments of this application will be clearly and completely described below in conjunction with the accompanying drawings in the embodiments of this application. Obviously, the described embodiments are only a part of the embodiments of this application, rather than all the embodiments. Based on the embodiments in this application, all other embodiments obtained by those skilled in the art without creative efforts belong to the scope of protection of this application.

[0085] In this application, the mention of "embodiment" means that the specific features, structures or characteristics described in connection with the embodiment can be included in at least one embodiment of this application. The phrase appears in various positions in the specification does not necessarily refer to the same embodiment, nor is it an independent or alternative embodiment mutually exclusive with other embodiments. Those skilled in the art explicitly and implicitly understand that the embodiments described in this application can be combined with other embodiments.

[0086] Embodiment 1

[0087] This embodiment discloses a method for predicting snoring during sleep based on physiological indicators of pulse wave in a waking state, as Figure 1 shown, and the specific steps are as follows:

[0088] S1. Construct a test scenario. The test device consists of a pulse wave signal measurement sensor and a supporting upper computer. Find non-solitary subjects to collect experimental data sets. The subject presses the index finger on the sensor. The entire test process includes five steps: sitting still, natural breathing, mouth breathing, left nasal cavity breathing, and right nasal cavity breathing; record the pulse wave signals during the processes of natural breathing, mouth breathing, left nasal cavity breathing, and right nasal cavity breathing, and each process lasts for T 1 , with an interval of rest for T 2 ; at the same time, ask the subject's roommate to monitor during the subject's sleep that night and inform the subject whether there is snoring phenomenon that night, and ask the subject whether he / she snored at night the next day.

[0089] Among them, the pulse wave acquisition sensor is specifically completed by using the MAX30102 sensor. The on-board wireless Bluetooth module HC-05 can be connected to the computer through Bluetooth, and the data can be uploaded through the upper computer. The subject relaxes the left hand, and the index finger naturally droops and fits the photoelectric sensor part of the sensor to collect data.

[0090] The minimum time requirement for short-time sampling based on the pulse wave is 120s. Therefore, for the detection duration T 1 of the pulse wave in each process in the experimental process, it is recommended to take the value of 120s, which has been verified to be available in actual experiments. In this embodiment, the time lengths T 1 of sitting still, natural breathing, mouth breathing, left nasal cavity breathing, and right nasal cavity breathing are all taken as 120s, and T 2The value of is 10s. The obtained pulse wave signals of the four processes are denoted as s raw (n), n = 0, 1, 2, 3, corresponding to four states of natural breathing, breathing only with the oral cavity, breathing only with the left nasal cavity, and breathing only with the right nasal cavity respectively.

[0091] S2. For the pulse wave signal s raw (n) collected in S1, n = 0, 1, 2, 3 are denoised respectively. The corresponding smoothed pulse wave signals are obtained by using low-pass filtering to remove high-frequency noise, corresponding to four states of natural breathing, breathing only with the oral cavity, breathing only with the left nasal cavity, and breathing only with the right nasal cavity respectively; then the corresponding pulse wave signals after baseline removal are obtained by the cubic spline interpolation baseline removal method, and then the cycle validity is evaluated, and the signal cycles with poor signal quality are removed to obtain the corresponding high-quality pulse wave signals. The specific process of the preprocessing of the pulse wave signal is as Figure 2 shown:

[0092] S201. According to the effective frequency range of the pulse wave, each segment of the pulse wave signal is subjected to low-pass filtering. The effective range of the pulse wave signal is 0 - 10Hz. Therefore, 10Hz low-pass filtering is used for signal filtering to remove high-frequency noise, and the smoothed pulse wave signal s smooth (n) is obtained, n = 0, 1, 2, 3, corresponding to four states of natural breathing, breathing only with the oral cavity, breathing only with the left nasal cavity, and breathing only with the right nasal cavity respectively;.

[0093] S202. Extract the starting point of each cycle of the signal. The local maximum extraction method is used to extract the starting point of each cycle. Based on the starting point, the front and back of each segment of the signal are truncated to avoid incomplete cycles that occur during sampling.

[0094] The starting point is obtained in the following way. For the starting point O i of its i-th cycle, there is where M is the number of sampling points within 1s. Due to the characteristics of signal sampling, there may be multiple points with equal values at the local maximum. Therefore, in the process of calculating the starting point, a temporary increment operation is performed on the middle point of multiple equal-valued local maximum points to ensure that there are no multiple equal-valued local maximum points during the calculation. The calculated local maximum point is the starting point of each cycle. The first and last cycles of the pulse wave signal are removed, and these cycles are often incomplete. After removal, the remaining number of cycles is K.

[0095] In the embodiment, the sampling rate of the sensor is 100Hz. Therefore, the number of samples within 1s is 100, rounded down, and the peak point is s smooth [O i >ssmooth [O i -d], d∈[-33,33].

[0096] In this embodiment, the sampling time T 1 is 120s. According to the common pulse rate in biology, the cycle number K is K∈[120,200].

[0097] S203, further, taking the starting point of each cycle as the base point, the signal is fitted using the cubic spline interpolation method, and the signal is subtracted to obtain the pulse wave signal after removing the baseline. For the pulse wave signal with K starting points, the cubic spline interpolation function is fitted. The starting point of each cycle (O i ,s smooth [O i ]), and the end of the last cycle (O K ,s smooth [K]) is used as the base point of cubic spline interpolation for fitting calculation, and the cubic spline fitting function can be obtained. On this basis, the signal baseline is removed, and the results are:

[0098]

[0099] In this embodiment, the end point of the last cycle will also be added with cubic spline interpolation fitting to remove the baseline of the last cycle.

[0100] S204, performing cycle validity evaluation on each signal cycle of the signal after removing the baseline, including Pearson coefficient calculation and skewness calculation, eliminating signal cycles whose Pearson coefficient and skewness do not meet the requirements, and obtaining the corresponding high-quality pulse wave signal s qualified (n), n=0,1,2,3, the specific processing process is as follows Figure 3 shown.

[0101] Specifically, based on the starting point O of the pulse wave signal i ,i=0,1,...,K-1, for signal s smooth The K-segment single-cycle pulse wave signal is obtained by period segmentation, where the i-th segment pulse wave signal is recorded as S i ,i=0,1,...,K-1.

[0102] For the period S i Perform skewness calculation and set the skewness threshold to θ s , where θ s is the average skewness of all periodic pulse waves. The pulse wave signal can be regarded as a right-skewed signal. In this embodiment, When the period S i Retained, with good performance.

[0103] For each single-cycle pulse wave signal screened by skewness, to avoid false alarms, it is necessary to eliminate invalid signals with skewness meeting the requirements. Therefore, the Pearson coefficient of adjacent cycles is calculated. Since the length of each cycle of the pulse wave signal is different, the number of sampling points in each cut cycle is inconsistent, and the Pearson correlation coefficient of adjacent cycles cannot be directly calculated. Therefore, cubic spline interpolation needs to be performed on the adjacent cycles obtained by cutting to obtain a signal X with the same number of sampling points. i and X i+1 , and then calculate the Pearson coefficient. Set the Pearson coefficient threshold to θ p . When P i > θ p , the cycle segment S i is retained. In this embodiment, the Pearson coefficient threshold θ p = 0.9.

[0104] S3. For high-quality pulse wave signals, extract the characteristic points of the pulse wave cycle. The characteristic points include the steepest rising point, the starting point, and the peak point. Based on the cycle characteristic points, extract the physiological characteristic data of the subject in this experiment. The physiological characteristic data includes heart rate HR, heart rate variability HRV, blood oxygen saturation SpO2, and the pulse wave morphology index MI. Based on the smoothed pulse wave signal, perform frequency domain analysis to obtain the respiratory rate RR. Using the pulse wave characteristics under natural breathing as the standard value, subtract the pulse wave characteristics during mouth breathing and single-nasal breathing respectively to obtain the physiological characteristic sequence. The specific process is as Figure 4 shown. The steps are as follows:

[0105] S301. For high-quality pulse wave signals after eliminating cycles with poor signal quality, extract the characteristic points of the pulse wave cycle. The characteristic points include the steepest rising point, the starting point, and the peak point. The above characteristic points are used for the extraction of physiological characteristics.

[0106] S302. Based on the characteristic points of the pulse wave cycle, calculate the heart rate, heart rate variability, and pulse wave morphology characteristics; based on the smoothed pulse wave signal, use frequency domain analysis and select the frequency value where the peak point in the frequency domain is located as the respiratory rate. The heart rate, heart rate variability, pulse wave morphology characteristics, and respiratory rate together constitute the pulse wave characteristic sequence. Finally, ppg_feature(n) = {HR(n), HRV(n), SpO2(n), MI(n), RR(n)}, n = 0, 1, 2, 3, corresponding to 4 states of natural breathing, breathing only with the mouth, breathing only with the left nasal cavity, and breathing only with the right nasal cavity respectively.

[0107] In this embodiment, the HRV indexes specifically include the low-frequency to high-frequency energy ratio (LF / HF), the root mean square of successive differences (RMSSD) of adjacent normal cardiac cycles, and the standard deviation of normal-to-normal intervals (SDNN). At the same time, HRV is approximately replaced by PRV.

[0108] In this embodiment, the pulse wave morphology index MI specifically includes the average upstroke time, the average downstroke time, the ratio of the upstroke time to the downstroke time, the ratio of the upstroke area to the downstroke area, and the volume pulse wave characteristic quantity.

[0109] S303. For the obtained pulse wave feature sequences ppg_featrue(n) in 4 states, where n = 0, 1, 2, 3, taking the pulse wave feature sequence ppg_featrue(0) under natural breathing as the standard value, subtracting the left nasal cavity pulse wave feature sequence from the natural breathing pulse wave feature sequence to obtain the left nasal cavity feature difference sequence Bias L , subtracting the right nasal cavity pulse wave feature sequence from the natural breathing pulse wave feature sequence to obtain the right nasal cavity feature difference sequence Bias R ; for the left nasal cavity feature difference sequence Bias L , and the right nasal cavity feature difference sequence Bias R , taking the maximum absolute value of each feature difference to form the single-nasal cavity feature difference sequence Bias N , using the pulse wave feature sequence under oral breathing and the natural breathing pulse wave feature sequence to perform a subtraction operation as the oral cavity feature difference sequence Bias M .

[0110] S4. Using the physiological feature sequence when the subject is awake as the input feature and whether the subject snores during sleep at night as the sample label, a training set is constructed.

[0111] S5. Using the training set to train the classifier, the inputs of the classifier include heart rate HR, heart rate variability HRV, blood oxygen saturation SpO2, pulse wave morphology index MI, and respiratory rate RR; the classifier maps the output data to a binary classification result through a combination of linear and non-linear methods; this result is the prediction of whether the sample snores and is also the output of the classifier;

[0112] The specific steps are as follows:

[0113] S501. For the data set obtained in step S4, feature selection is performed during the actual training process to eliminate features with less correlation and form an optimal subset of feature data.

[0114] In this embodiment, first, use the original dataset to train a random forest model once, and based on the ability of the features to reduce the classification information entropy, statistically calculate the importance of each feature for the random forest model and sort them. The importance of the original dataset for the random forest model is as Figure 5 shown. There are a total of 22 features. Retain the top 50% (11 features) of the sorted features to form an optimal feature data subset.

[0115] S502. Split the optimal feature data subset into a training set and a test set. Use the training set to train the classifier and the test set to verify the performance of the classifier.

[0116] In this embodiment, the classifier uses a random forest, where the number of decision trees is set to 500, and the training criterion adopts the principle of minimizing information entropy.

[0117] The importance of the sample features in the optimal feature data subset for the random forest model is as Figure 6 shown. It can be seen from the figure that the sample features in the optimal feature data subset have relatively high importance for random forest classification, indicating that feature selection based on random forest importance can effectively select important features in classification.

[0118] S503. In this embodiment, step S502 is repeated in a 5-fold cross-validation manner to obtain the average performance evaluation result of the model.

[0119] In this embodiment, the accuracy of the snoring prediction and recognition model is statistically analyzed, and its average performance evaluation is shown in Table 1 below:

[0120] Table 1. Snoring prediction and recognition accuracy table in Embodiment 1

[0121]

[0122] S6. Input the input features of the new user into the classifier trained in S5 to obtain the prediction result of whether the new user snores.

[0123] S601. For the new user, repeat steps S1 to S4 to obtain the input features of the user Feature = {Bias N , Bias M}. Different from the subjects during dataset collection, at this time, it is no longer required that the user knows whether they snore at night.

[0124] S602. Obtain the input features of the user through steps S2 to S4;

[0125] S603. Input the input features of the user into the snoring prediction classifier to predict whether the user has a snoring problem at night.

[0126] In summary, using the physiological feature sequence of the subject when awake as the input feature and whether snoring occurs or not as the sample label, the construction of the snoring prediction classifier is completed. The snoring prediction classifier has a high accuracy rate and is used for the subsequent actual application of predicting snoring of new users at night.

[0127] Embodiment 2

[0128] This embodiment discloses a method for predicting snoring during sleep based on physiological indicators of pulse waves in a waking state, including the following steps.

[0129] S1. Refer to the corresponding steps in Embodiment 1, which will not be elaborated here. At this time, for a new user whose snoring status is unknown, different from the subjects during data set collection, it is no longer required that the user knows whether they snore at night.

[0130] S2. Refer to the corresponding steps in Embodiment 1, which will not be elaborated here. Among them, the cycle screening effect is as Figure 7 , Figure 8 shown, Figure 7 is the signal before cycle screening, Figure 8 is the signal after cycle screening. The low-quality signal pulse wave cycles in the signal before cycle screening are as Figure 7 shown and are removed after being screened by the tandem screening of skewness - adjacent Pearson coefficient. At the same time, all valid signal cycles are retained, as Figure 8 shown. It can be seen that the cycle screening method based on the tandem of skewness - adjacent Pearson coefficient has a good screening effect.

[0131] S3. Refer to the corresponding steps in Embodiment 1, which will not be elaborated here.

[0132] S4. Refer to the corresponding steps in Embodiment 1, which will not be elaborated here. However, at this time, no data set is constructed, and the input features of the user are only used as the input of the already trained snoring prediction model.

[0133] S6. Input the input features of the new user into the snoring prediction model to obtain the prediction result of whether the user snores.

[0134] In summary, according to the snoring prediction classifier, input the physiological feature sequence of the new user when awake to obtain the snoring prediction result, thereby completing the actual application of snoring prediction based on the physiological indicators of the pulse wave of the new user in the waking state.

[0135] The technical features of the above embodiments can be combined arbitrarily. For the sake of brief description, not all possible combinations of the technical features in the above embodiments are described. However, as long as the combinations of these technical features do not conflict, they should all be considered as the scope recorded in this specification.

[0136] The above embodiments are the preferred embodiments of the present invention, but the embodiments of the present invention are not limited to the above embodiments. Any other changes, modifications, substitutions, combinations, and simplifications made without departing from the spirit and principle of the present invention shall be equivalent replacement methods and are all included in the protection scope of the present invention.

Claims

1. A method for predicting snoring in a sleeping state based on pulse wave physiological indicators in a waking state, characterized in that: The sleep state snoring prediction method comprises the following steps: S1. Construct a test scenario. The test equipment consists of a pulse wave signal measurement sensor and a matching host computer. Find subjects who do not live alone to collect experimental data sets. Place the subject's index finger on the sensor. The entire test process is as follows: sit quietly, breathe naturally, breathe through the mouth, breathe through the left nasal cavity, and breathe through the right nasal cavity. Record the pulse wave signals during the natural breathing, mouth breathing, left nasal cavity breathing, and right nasal cavity breathing processes. At the same time, require the subject to have his roommate monitor the subject while he is sleeping that night, and inform the subject whether he snores that night. Ask the subject whether he snores at night the next day. S2, denoising each pulse wave signal collected in step S1, using low-pass filtering to remove high-frequency noise to obtain a smooth pulse wave signal, then using a cubic spline interpolation baseline removal method to obtain a pulse wave signal after baseline removal, and then performing a cycle validity evaluation to remove signal cycles with poor signal quality to obtain a high-quality pulse wave signal; S3. Extract pulse wave period feature points from high-quality pulse wave signals. Feature points include the fastest rising point, starting point, peak point and period-based feature points. Extract the subject's physiological feature data in this experiment. Physiological feature data include heart rate HR, heart rate variability HRV, blood oxygen saturation SpO2 and pulse wave morphology index MI. Perform frequency domain analysis on the smoothed pulse wave signal to obtain respiratory rate RR. Take the pulse wave feature under natural breathing as the standard value, use mouth breathing and single nasal breathing to make a difference, and obtain the physiological feature sequence. S4. Use the subject's physiological feature sequence when awake as input features and whether he or she snores during sleep at night as sample labels to construct a training set; S5. Use the training set to train the classifier, the input of which includes heart rate HR, heart rate variability HRV, blood oxygen saturation SpO2, pulse wave morphology index MI, and respiratory rate RR; the classifier maps the output data to a binary classification result by combining linear and nonlinear methods; the result is the prediction of whether the sample is snoring, which is also the output of the classifier; S6. For a new user who does not know whether he or she snores, when the user is awake, steps S1 to S4 are repeated to obtain test data and input the data into the classifier to predict whether the new user snores during sleep.

2. The method for predicting snoring in a sleeping state based on pulse wave physiological indicators in a waking state according to claim 1, characterized in that: The process of step S1 is as follows: S101. Construct a test scenario. Under conditions of suitable temperature and ventilation, find subjects who do not live alone to collect experimental data sets. The subjects press their fingers on the pulse wave signal sensor and sit with their backs against a chair. S102, asking the subject to sit quietly for a duration of T1, so that the subject can calm down and become familiar with the test environment, and avoid interference of the test environment on the physiological signal; S103, after the interval T2 allows the fingers to be fully relaxed, the subject's natural breathing lasts for T1, and the pulse wave data is collected by the pulse wave signal sensor as the reference signal s raw (0); S104, after the interval T2 allows the fingers to be fully relaxed, the subject breathes only through the mouth, for a duration of T1, and collects pulse wave data to obtain a pulse wave signal s raw (1) To assess oral ventilation; S105, after the interval T2 allows the fingers to be fully relaxed, the subject breathes only through the left nostril for a duration of T1, collects pulse wave data, and obtains a pulse wave signal s raw (2) to assess ventilation in a single nasal cavity; S106, after the interval T2 allows the fingers to be fully relaxed, the subject breathes only through the right nostril for a duration of T1, collects pulse wave data, and obtains a pulse wave signal s raw (3) to assess ventilation of a single nasal cavity; The above pulse wave signal is recorded as s raw (n), n = 0, 1, 2, 3, corresponding to the four states of natural breathing, breathing only through the mouth, breathing only through the left nasal cavity, and breathing only through the right nasal cavity; S107. Ask the subject to have his roommate monitor his sleep to find out whether he snores that night. Ask the subject whether he snores the next day, and use the label Label as the training set sample for supervised learning, where Label∈{0,1}.

3. The method for predicting snoring in a sleeping state based on pulse wave physiological indicators in a waking state according to claim 1, characterized in that: The process of step S2 is as follows: S201. According to the effective frequency range of the pulse wave, the pulse wave signal is recorded as s raw (n), n=0,1,2,3 respectively perform low-pass filtering to remove high-frequency noise and obtain the corresponding smoothed pulse wave signal s smooth (n), n = 0, 1, 2, 3, corresponding to the four states of natural breathing, breathing only through the mouth, breathing only through the left nasal cavity, and breathing only through the right nasal cavity; S202, extracting the starting point of each cycle of the signal, using the local maximum point extraction method to extract the starting point of each cycle, based on the starting point, truncating the front and back of each section of the signal, removing the starting cycle of the signal and the last cycle, to avoid interference with the signal caused by incomplete cycles that occur during sampling before and after; S203, taking the starting point of each cycle of the smoothed pulse wave as the base point, using the cubic spline interpolation method to fit the signal, subtracting the signal, and obtaining the pulse wave signal s after removing the baseline detrend ; S204, performing cycle validity evaluation on each signal cycle of the signal after removing the baseline, including skewness calculation and Pearson coefficient calculation, eliminating signal cycles whose Pearson coefficient and skewness do not meet the requirements, and obtaining the corresponding high-quality pulse wave signal s qualified (n), n = 0, 1, 2, 3, corresponding to four states: natural breathing, breathing using only the mouth, breathing using only the left nostril, and breathing using only the right nostril.

4. The method for predicting snoring in a sleeping state based on pulse wave physiological indicators in a waking state according to claim 3, characterized in that: The process of step S3 is as follows: S301, for the high-quality pulse wave signal after removing the period with poor signal quality, extract the pulse wave period feature points, the feature points include the fastest rising point, the starting point, and the peak point, and the above feature points are used for extracting physiological characteristics; S302, based on the pulse wave period feature points, calculate the heart rate, heart rate variability, and pulse wave morphological features; based on the smoothed pulse wave signal, use frequency domain analysis to select the frequency value where the frequency domain peak point is located as the respiratory rate; the heart rate, heart rate variability, pulse wave morphological features, and respiratory rate together constitute the pulse wave feature sequence, and finally obtain the pulse wave feature sequence ppg_feature(n) = {HR(n), HRV(n), SpO2(n), MI(n), RR(n)}, n = 0, 1, 2, 3, corresponding to four states: natural breathing, breathing only using the mouth, breathing only using the left nasal cavity, and breathing only using the right nasal cavity; S303, for the pulse wave feature sequences ppg_featrue(n) of the four states obtained, n=0, 1, 2, 3, take the pulse wave feature sequence ppg_featrue(0) under natural breathing as the standard value, and use the left nasal pulse wave feature sequence to make a difference with the natural breathing pulse wave feature sequence to obtain the left nasal feature difference sequence Bias L , use the right nasal pulse wave feature sequence and the natural breathing pulse wave feature sequence to obtain the right nasal feature difference sequence Bias R ; Left nasal feature difference sequence Bias L , right nasal feature difference sequence Bias R Each feature difference takes the largest absolute value to form a single nose feature difference sequence Bias N , use the pulse wave feature sequence under oral breathing and the natural breathing pulse wave feature sequence to perform difference calculation as the oral feature difference sequence Bias M .

5. The method for predicting snoring in a sleeping state based on pulse wave physiological indicators in a waking state according to claim 1, characterized in that: In step S4, the single nose feature difference sequence and the oral feature difference sequence are concatenated to obtain sample input features as a training set, and whether the known subject snores is used as a sample label to construct a data set. The process is as follows: S401. For a snoring subject, the sample composition is: Input feature Feature = {Bias N ,Bias M }, sample label Label = {1}; S402: For a non-snoring subject, the sample composition is: Input feature Feature = {Bias N ,Bias M }, sample label Label = {0}.

6. The method for predicting snoring in a sleeping state based on pulse wave physiological indicators in a waking state according to claim 1, characterized in that: The process of step S5 is as follows: S501, for the data set obtained in step S4, feature selection is performed in the actual training process to eliminate features with less correlation and form an optimal feature data subset; S502, splitting the optimal feature data subset into a training set and a test set, using the training set to train the classifier, and using the test set to test the performance of the classifier; S503: Repeat step S502 using a K-fold cross-validation method to obtain an average performance evaluation result of the classifier.

7. The method for predicting snoring in a sleeping state based on pulse wave physiological indicators in a waking state according to claim 1, characterized in that: The process of step S6 is as follows: S601, for a new user who does not know whether he or she snores, repeat step S1 to collect the user's pulse wave physiological data. Unlike the subject, the user is no longer required to know whether he or she snores at night; S602, obtaining the input features of the user through steps S2 to S4; S603: Input the input features of the user into a snoring prediction classifier to predict whether the user has a snoring problem at night.

Images