Obstructive sleep apnea detection methods, systems, devices, and storage media

Abstract

The application provides a method, system, device and storage medium for detecting obstructive sleep apnea, and relates to the technical field of signal processing. The method comprises the following steps: acquiring pressure signals detected by at least two bed-mounted pressure sensors in a human sleep state, and acquiring a crowd type feature and a pre-sleep activity feature of a person to be detected; respectively performing quality assessment on each pressure signal, and determining a target pressure signal based on the quality assessment result; performing pretreatment on the target pressure signal, and performing feature extraction on the pretreated target pressure signal in a time domain and a frequency domain; inputting the extracted multi-dimensional features, the crowd type feature and the pre-sleep activity feature of the person to be detected into a pre-trained machine learning model to obtain a classification result of whether the person to be detected has obstructive sleep apnea. The application realizes accurate detection of OSA without using a PSG device.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of signal processing, in particular to a method and system for detecting obstructive sleep apnea, a device and a storage medium. BACKGROUND

[0002] Sleep disordered breathing, especially obstructive sleep apnea (OSA), is a common and serious sleep disease that affects health. In the sleep process, the upper airway of OSA patients will repeatedly obstruct, causing apnea or significant reduction in ventilation, and thus causing a series of problems such as hypoxemia and sleep structure disorder. Long-term OSA can increase the risk of many chronic diseases such as hypertension, coronary heart disease, arrhythmia, cerebrovascular disease, and diabetes, and seriously affect the quality of life and life of patients.

[0003] Traditional OSA diagnosis mainly relies on polysomnography (PSG). PSG can comprehensively and accurately monitor various physiological parameters during sleep, such as electroencephalogram, electrooculogram, electromyogram, electrocardiogram, respiratory airflow, chest and abdominal respiratory movement, and blood oxygen saturation, thereby providing a reliable basis for OSA diagnosis and condition assessment. However, this method has many limitations. PSG requires patients to sleep in a hospital sleep monitoring room, the monitoring equipment is complex, and the patient needs to wear many electrodes and sensors, which not only brings great inconvenience and discomfort to the patient, affects the sleep quality, and may cause deviation in the monitoring results, but also has high monitoring costs, and requires professional technical personnel to operate and analyze, making it difficult for PSG to be widely used for daily screening and long-term home monitoring. SUMMARY

[0004] Therefore, the present application provides a method and system for detecting obstructive sleep apnea, a device and a storage medium.

[0005] In a first aspect, the present application provides a method for detecting obstructive sleep apnea, comprising:

[0006] obtaining pressure signals detected by at least two bed-mounted pressure sensors in a sleep state of a human body, and crowd type characteristics and pre-sleep activity characteristics of a person to be detected;

[0007] respectively performing quality assessment on each pressure signal, and determining a target pressure signal based on the quality assessment result;

[0008] preprocessing the target pressure signal, and performing feature extraction on the preprocessed target pressure signal in time domain and frequency domain;

[0009] inputting the extracted multi-dimensional features, the crowd type characteristics and the pre-sleep activity characteristics of the person to be detected into a pre-trained machine learning model to obtain a classification result of whether the person to be detected has obstructive sleep apnea.

[0010] In an embodiment, the machine learning model is a Bayesian model; the extracted multi-dimensional features, the crowd type feature of the to-be-detected person, and the pre-sleep activity feature are input into the pre-trained machine learning model to obtain a classification result of whether the to-be-detected person has obstructive sleep apnea, including:

[0011] The extracted multi-dimensional features, the crowd type feature of the to-be-detected person, and the pre-sleep activity feature are input into the Bayesian model, so that the Bayesian model calls a crowd-specific feature parameter library according to the crowd type feature of the to-be-detected person;

[0012] Obtain the crowd type weight coefficient and the activity weight coefficient corresponding to the crowd type feature and the pre-sleep activity feature of the to-be-detected person, respectively;

[0013] Adjust the weights of the time domain features and the frequency domain features based on the crowd type weight coefficient and the activity weight coefficient, and determine the posterior probability of the two results of whether the to-be-detected person has obstructive sleep apnea in combination with the called crowd-specific parameters;

[0014] Determine the classification result of whether the to-be-detected person has obstructive sleep apnea according to the posterior probability of the two results.

[0015] In an embodiment, the adjustment of the weights of the time domain features and the frequency domain features based on the crowd type weight coefficient and the activity weight coefficient, and the determination of the posterior probability of the two results of whether the to-be-detected person has obstructive sleep apnea in combination with the called crowd-specific parameters, include:

[0016] Determine the calibration weight of each time domain feature and each frequency domain feature according to the crowd type weight coefficient and the activity weight coefficient;

[0017] According to the crowd type of the to-be-detected person, call the feature conditional probability distribution of the corresponding crowd from the corresponding crowd-specific feature parameter library;

[0018] Determine the posterior probability of the two results of whether the to-be-detected person has obstructive sleep apnea according to the extracted multi-dimensional features, the calibration weight of each time domain feature and each frequency domain feature, and the feature conditional probability distribution.

[0019] In an embodiment, the quality assessment of each channel pressure signal, and the determination of the target pressure signal based on the quality assessment result, include:

[0020] Calculate the energy and the standard deviation of the sub-signal of each channel signal in the target frequency band to obtain the quality assessment result in the target frequency band;

[0021] The fusion weights of each signal are determined based on the frequency band quality assessment results, and the sub-signals of each signal in the same frequency band are weighted and fused based on the fusion weights of each signal to obtain the target pressure signal.

[0022] In one embodiment, the step of performing quality assessments on each pressure signal and determining the target pressure signal based on the quality assessment results includes:

[0023] Perform a sliding window analysis for each signal with a preset time, and calculate the signal energy and standard deviation within the window;

[0024] The corresponding energy fluctuation coefficient is determined based on the signal energy within the window of each signal.

[0025] The optimal signal is determined based on the energy fluctuation coefficient and standard deviation of each signal, and the optimal signal is used as the target pressure signal.

[0026] In one embodiment, the preprocessing of the target pressure signal and the extraction of time-domain and frequency-domain features from the preprocessed target pressure signal include:

[0027] Low-frequency noise below a predetermined frequency is removed by adaptive Kalman filtering, and multiple frequency band signals are obtained by multi-layer wavelet decomposition, wherein the multiple frequency band signals include respiratory main frequency band signals and cardiovascular main frequency band signals.

[0028] The respiratory main frequency band signal is detected by a peak detection algorithm to identify respiratory peaks. A respiratory threshold is determined based on the signal mean of the respiratory main frequency band signal. The respiratory frequency and respiratory cycle variation coefficient are determined according to the identified respiratory peaks and the respiratory threshold. The maximum amplitude, minimum amplitude and amplitude ratio of the respiratory signal are extracted from the respiratory main frequency band signal.

[0029] Wavelet denoising is performed on the cardiovascular main frequency band signal to extract the heartbeat signal, and the heart rate variability index is calculated based on the heartbeat signal and a preset time window;

[0030] Fast Fourier Transform is performed on time period signals selected from the respiratory dominant frequency band signal and the cardiovascular dominant frequency band signal to calculate the energy proportion of the respiratory dominant frequency band and the energy proportion of the cardiovascular dominant frequency band, respectively.

[0031] Secondly, this application also provides an obstructive sleep apnea detection system, including: a processing module and at least two bed-mounted pressure sensors;

[0032] Each bed pressure sensor is designed to be installed on the mattress at the position corresponding to the human chest.

[0033] The processing module is connected with each bed load pressure sensor respectively, and is used for executing the obstructive sleep apnea detection method of the first aspect.

[0034] In an embodiment, the processing module comprises a signal conditioning module, a low-pass filter, a high-pass filter, a data acquisition module and a data processing module.

[0035] The signal conditioning module is used for receiving pressure signals output by each bed load pressure sensor, and amplifying each pressure signal to obtain an amplified pressure signal.

[0036] The low-pass filter is used for filtering high-frequency noise of the amplified pressure signal.

[0037] The high-pass filter is used for filtering the signal output by the low-pass filter to suppress baseline drift.

[0038] The data acquisition module is used for performing analog-to-digital conversion on the signal output by the high-pass filter to generate a digital signal.

[0039] The data processing module is used for the digital signal output by the data acquisition module, and executes the obstructive sleep apnea detection method of the first aspect.

[0040] In a third aspect, the application further provides an electronic device comprising a processor and a memory; the memory stores a computer program, wherein the computer program, when executed by the processor, implements the obstructive sleep apnea detection method of the first aspect.

[0041] In a fourth aspect, the application further provides a computer storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements the obstructive sleep apnea detection method of the first aspect.

[0042] The obstructive sleep apnea detection method of the application has the following beneficial effects compared with related art:

[0043] 1. The obstructive sleep apnea detection method of the present application, by acquiring at least two pressure signals detected by the bed-mounted pressure sensor in the human sleep state, and respectively evaluating the quality of each pressure signal, determining the target pressure signal based on the quality evaluation result, so as to obtain the target pressure signal with better quality. After obtaining the target pressure signal, the target pressure signal can be preprocessed, and then the time domain and frequency domain features of the preprocessed target pressure signal are extracted, and the multi-dimensional features related to OSA are accurately extracted from the pressure signal. On this basis, considering the influence of the crowd type characteristics and the pre-sleep activity characteristics of the to-be-detected person, the extracted multi-dimensional features, the crowd type characteristics and the pre-sleep activity characteristics of the to-be-detected person are input into the pre-trained machine learning model, and the multi-dimensional physiological characteristics are used to improve the diagnostic accuracy; at the same time, the model can break through the limitation of single crowd standard by relying on the crowd type characteristics, optimize the judgment logic for high-risk crowds and pre-sleep activities, and enhance the individualized adaptation ability, and further improve the classification accuracy of OSA.

[0044] 2. The obstructive sleep apnea detection method of the present application does not need to rely on complex and expensive PSG equipment, at least two bed-mounted pressure sensors are arranged on the mattress, and signals are collected by the bed-mounted pressure sensor, which is suitable for primary medical care and home screening scenes, and has high convenience.

[0045] 3. Based on the crowd type weight coefficient and the activity weight coefficient, the weights of the time domain features and the frequency domain features related to respiratory stability can be adjusted, and the weights of the time domain features and the frequency domain features related to respiratory stability can be adjusted. Subsequently, the pre-constructed crowd-specific parameters are called, the time domain / frequency domain features with adjusted weights are compared with these parameters, and the posterior probability is calculated by combining the Bayes formula. Based on the OSA prior probability of this crowd, the likelihood is calculated by using the physiological characteristics with adjusted weights, and the influence of the crowd / activity weight is integrated, and finally the posterior probability of the two results of "suffering from OSA" and "not suffering from OSA" is output, and finally the classification result of whether the to-be-detected person suffers from obstructive sleep apnea is determined according to the posterior probability of the two results. The above classification process takes the probability value as the core basis, combines the crowd risk characteristics, balances the risk of missed judgment and misjudgment, and ensures the accuracy of the classification result. BRIEF DESCRIPTION OF DRAWINGS

[0046] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the related art, the drawings needed to be used in the embodiment or related art description will be briefly introduced. Obviously, the drawings in the following description are only some embodiments of the present application, and those skilled in the art can obtain other drawings according to these drawings without creating any creative labor.

[0047] Figure 1A flowchart of a method for detecting obstructive sleep apnea in an embodiment of the present application is shown.

[0048] Figure 2 A flowchart of step S104 in an embodiment of the present application is shown.

[0049] Figure 3 A flowchart of step S103 in an embodiment of the present application is shown.

[0050] Figure 4 A structural diagram of a system for detecting obstructive sleep apnea in an embodiment of the present application is shown. DETAILED DESCRIPTION

[0051] The technical solutions in the embodiments of the present application will be described clearly and completely below in conjunction with the embodiments of the present application. Obviously, the described embodiments are only some of the embodiments of the present application, but not all the embodiments of the present application. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative work fall within the scope of protection of the present application.

[0052] In some embodiments, as shown in FIG. 1, the present application provides a method for detecting obstructive sleep apnea, which comprises the following steps S101-S104. Figure 1

[0053] S101: Obtain pressure signals detected by at least two bed-mounted pressure sensors in a human sleep state, and population type characteristics and pre-sleep activity characteristics of a person to be detected.

[0054] The bed-mounted pressure sensors can be deployed in a built-in mattress or a bed surface laying manner. The bed-mounted pressure sensors are started before the person to be detected sleeps, and continuously collect signals of a complete sleep cycle, converting small pressure changes in human sleep into electrical signals. The population type characteristics of the person to be detected focus on high-risk population types and normal population types, which can cover age, body weight and size, and underlying diseases, etc. For example, the population can be divided into elderly population, overweight population, and normal population. The pre-sleep activity characteristics are directed to non-pathological factors that can interfere with OSA diagnosis within a certain time before sleep, which can include diet, drink, and physical activity, etc. For example, drinking and pre-sleep vigorous exercise.

[0055] S102: Perform quality assessment on each pressure signal respectively, and determine a target pressure signal based on the quality assessment result.

[0056] ​In the application, the corresponding feature parameters can be calculated based on the pressure signals, and then the quality of each pressure signal can be evaluated based on a preset standard. The signal with better quality can be selected as the target pressure signal based on the quality evaluation result, or the signal fusion processing can be performed on each pressure signal based on the quality evaluation result, and the fused signal can be taken as the target pressure signal. The standard for quality evaluation can consider the anti-interference ability, that is, the fluctuation parameters of each pressure signal are evaluated. The standard for quality evaluation can also consider the signal recognition degree, and the signal quality is evaluated by judging whether the signal can clearly distinguish the normal physiological action in sleep from the interference signal.

[0057] S103: The target pressure signal is preprocessed, and the time domain and frequency domain features of the preprocessed target pressure signal are extracted.

[0058] Preprocessing can include denoising processing. For example, signal noise can be removed by wavelet filtering or mean filtering, and then a small amount of breakpoints caused by temporary signal loss can be corrected by interpolation method to ensure signal continuity.

[0059] It can be understood that the time domain feature focuses on the intuitive performance of the signal in the time dimension, and the frequency domain feature needs to be converted from the time domain signal to the frequency domain through Fourier transform. The frequency characteristics of the signal can be mined from the frequency dimension to obtain hidden information that cannot be intuitively presented in the time domain signal, and provide supplementary basis for sleep state evaluation. The two types of features jointly cover the time law and frequency characteristics of the signal, and provide quantitative data support for subsequent sleep stage analysis and abnormal behavior recognition, thereby facilitating improvement of the classification accuracy of OSA.

[0060] S104: The extracted multi-dimensional features, the crowd type features of the to-be-detected person, and the pre-sleep activity features are input into a pre-trained machine learning model to obtain a classification result of whether the person has obstructive sleep apnea.

[0061] The machine learning model of the application can adopt a random forest model, a support vector machine, a Bayesian model, or the like.

[0062] In the application, sample data of OSA patients diagnosed by PSG and healthy people can be collected, and features can be extracted to construct a training set.

[0063] It can be understood that the time domain features and the frequency domain features extracted from the detected pressure signals can directly reflect physiological signals related to OSA, such as changes in respiratory airflow during sleep, body posture fluctuations, and the like. The population type features and the pre-sleep activity features of the to-be-detected person are closely related to the OSA incidence risk, and the population type features and the pre-sleep activity features of the to-be-detected person can assist the model to adapt to individual tendencies. Therefore, after the three types of features are jointly input into the pre-trained machine learning model, the machine learning model can first associate and match the features, and then calculate the classification probability through an internal algorithm, and finally output the classification result of “suffering from OSA” or “not suffering from OSA”, realizing the non-invasive and convenient detection of OSA.

[0064] It should be noted that in the application, the model can also be updated based on newly collected data, and the update is started after each new sample of a preset number (for example, 50) is collected. The newly collected sample data can be filtered first to ensure signal integrity (effective fragment ratio > 80%, and body motion interference > 30% of the fragment is removed) and label reliability (new samples are reviewed by PSG, AHI error < 5).

[0065] The model update can be triggered based on feature drift detection, specifically: KS test (significance level 0.05) is performed on the multi-dimensional features (for example, 12-dimensional features) of the new sample and the original training set respectively, when ≥3-dimensional features pass the test, full update is triggered; otherwise, incremental update is performed.

[0066] The model update strategy can include full update and incremental update.

[0067] Full update: merge new samples and original data set, retrain model. Use adaptive bandwidth Gaussian kernel density estimation, bandwidth is dynamically adjusted according to feature variance 1 / 5, Laplace smoothing coefficient is optimized to .

[0068] Incremental update: only use new samples to update model parameters, continuously update mean and variance of features, and the corresponding formula is as follows:

[0069]

[0070]

[0071] wherein, represents the mean of the updated continuous features, represents the variance of the continuous features before updating, represents the variance of the continuous features in the new sample set; represents the variance of the updated continuous features, represents the variance of the continuous features before updating, represents the variance of the continuous features in the new sample set.

[0072] Discrete feature update conditional probability

[0073]

[0074] denotes the conditional probability of the discrete feature taking x given that the class is C after the update; denotes the number of samples in the class C before the update; denotes the conditional probability of the discrete feature taking x given that the class is C before the update; denotes the number of samples in the class C in the new sample set; denotes the conditional probability of the discrete feature taking x given that the class is C in the new sample set.

[0075] The OSA detection method provided by the present application obtains at least two pressure signals detected by bed-mounted pressure sensors in a human sleep state, respectively evaluates the quality of each pressure signal, determines a target pressure signal based on the quality evaluation result, and thus obtains a target pressure signal with better quality. After the target pressure signal is obtained, the target pressure signal can be preprocessed, and then the preprocessed target pressure signal is subjected to feature extraction in the time domain and the frequency domain, so as to accurately extract multi-dimensional features related to OSA from the pressure signal. On this basis, the extracted multi-dimensional features, the crowd type feature and the pre-sleep activity feature of the to-be-detected person are input into a pre-trained machine learning model, the diagnostic accuracy is improved through multi-dimensional physiological features; at the same time, the model can break through the limitation of a single crowd standard by relying on the crowd type feature, optimize the judgment logic for high-risk crowds and pre-sleep activities of the person, and enhance the individualized adaptation capability, thereby further improving the classification accuracy of OSA. At the same time, the OSA detection method provided by the present application does not need to rely on a complex and expensive PSG device, at least two bed-mounted pressure sensors are arranged on a mattress, and signals can be collected through the bed-mounted pressure sensors, so that the method is suitable for primary medical treatment and home screening scenes and has high convenience.

[0076] In some embodiments, the machine learning model is a Bayesian model. As shown in Figure 2 In step S104, the extracted multi-dimensional features, the crowd type feature and the pre-sleep activity feature of the to-be-detected person are input into a pre-trained machine learning model to obtain a classification result of whether the person has OSA, including the following steps S201 to S204.

[0077] S201: The extracted multi-dimensional features, the crowd type feature and the pre-sleep activity feature of the to-be-detected person are input into the Bayesian model, so that the Bayesian model calls a crowd-specific feature parameter library according to the crowd type feature of the to-be-detected person.

[0078] Among them, the population-specific characteristic parameter library pre-stores OSA-related specific parameters for different population types.

[0079] Understandably, after inputting the extracted multi-dimensional features, the population type features of the individuals to be tested, and the pre-sleep activity features into the Bayesian model, the Bayesian model will analyze the population type features of the individuals to be tested. Based on these features, it will automatically call upon a pre-built population-specific feature parameter library to determine the prior probability of the corresponding population type of the individuals to be tested. Then, the Bayesian model can use the population-specific parameters to perform probabilistic inference on the multi-dimensional features and pre-sleep activity features, avoiding the neglect of physiological differences between different populations due to the use of general parameters, thereby improving the accuracy of subsequent OSA detection.

[0080] S202: Obtain the population type weight coefficient and activity weight coefficient corresponding to the population type characteristics and bedtime activity characteristics of the person to be tested.

[0081] In applications, algorithms such as logistic regression and random forest can be used to calculate the contribution of population type characteristics and pre-sleep activity characteristics to OSA based on historical sample data. For example, overweight (BMI ≥ 30) is much more associated with OSA than normal weight (BMI 18.5-23.9), and drinking alcohol before bed significantly increases the risk of OSA due to relaxation of airway muscles, with a much higher association than using electronic devices before bed. Therefore, the weight coefficients of corresponding features will be different, with higher contribution resulting in larger weight coefficients, ensuring that the impact priority of different population characteristics on OSA detection is accurately reflected.

[0082] For example, a weighting coefficient ω can be introduced when calculating the posterior probability. For instance, if an overweight patient has a BMI > 30, the initial weight of the "breathing amplitude ratio" feature is increased to 1.4 times; if alcohol is detected at dinner that day, the weight of the "nighttime blood oxygen fluctuation" feature is increased by 1.3 times in the nighttime detection.

[0083] S203: Adjust the weights of time-domain and frequency-domain features based on population type weight coefficients and activity weight coefficients, and combine them with the population-specific parameters to determine the posterior probability of whether the person being tested has obstructive sleep apnea.

[0084] It can be understood that the core logic of adjusting the time domain and frequency domain feature weights is to adapt the influence degree of the physiological signal features to the individual attributes and short-term behaviors of the to-be-detected person. Based on the population type weight coefficient and the activity weight coefficient, the weights of the time domain features and the frequency domain features related to the respiratory stability can be adjusted, and the weights of the time domain features and the frequency domain features related to the respiratory stability can be adjusted. Then, the pre-constructed population-specific parameters are called, the time domain / frequency domain features with the adjusted weights are compared with these parameters, and the posterior probability is calculated based on the Bayes formula. Based on the OSA prior probability of the population, the likelihood of the physiological features with the adjusted weights is calculated, and the influence of the population / activity weight is integrated, and finally the posterior probability of the two results of "having OSA" and "not having OSA" is output.

[0085] S204: Determine whether the to-be-detected person has obstructive sleep apnea according to the posterior probability of the two results.

[0086] It can be understood that after the posterior probability of the two results of having OSA and not having OSA, three cases can be judged: the first case is that the posterior probability of having OSA is significantly higher than the preset threshold, and is much larger than the probability of not having OSA, for example, the probability of having OSA is 0.89, and the probability of not having OSA is 0.11, 0.89>0.5 (common threshold), and the result is directly determined as having obstructive sleep apnea; the second case is that the posterior probability of not having OSA is higher than the threshold, and the value is dominant, for example, for a young normal weight female, the posterior probability of not having OSA is 0.72, and the posterior probability of having OSA is 0.28, 0.72>0.5, which can be determined as not having. The above classification process takes the probability value as the core basis, combines the population risk characteristics, balances the risk of missed judgment and misjudgment, and ensures the accuracy of the classification result.

[0087] In some embodiments, in step S203, based on the population type weight coefficient and the activity weight coefficient, the weights of the time domain features and the frequency domain features are adjusted, and the posterior probability of the two results of whether the to-be-detected person has obstructive sleep apnea is determined in combination with the called population-specific parameters, including: determining the calibration weights of each time domain feature and each frequency domain feature according to the population type weight coefficient and the activity weight coefficient; calling the feature conditional probability distribution of the corresponding population from the corresponding population-specific feature parameter library according to the population type of the to-be-detected person; determining the posterior probability of the two results of whether the to-be-detected person has obstructive sleep apnea according to the extracted multi-dimensional features, the calibration weights of each time domain feature and each frequency domain feature, and the feature conditional probability distribution.

[0088] In the application, for each population group, the feature conditional probability distribution of its PSG samples can be stored separately; at the same time, the population type weight coefficient for that population group can also be stored. For example, in the parameter library for overweight patients, the conditional probability of abnormal respiratory amplitude ratio is stored, and the population type weight coefficient corresponding to this feature is determined. For example, the population type weight coefficient for overweight patients is 1.4; while the population type weight coefficient for normal adults is 1.0.

[0089] Let X be the feature set of the person to be tested (including multi-dimensional physiological characteristics, population type characteristics, and pre-sleep activity characteristics). Then, P(C1|X) represents the posterior probability that the person to be tested has OSA given the feature X, and P(C2|X) represents the posterior probability that the person to be tested does not have OSA given the feature X. These two probabilities are calculated by a pre-trained machine learning model, and their sum is 1.

[0090] Combining the posterior probability formula of Naive Bayes, the weight coefficient ω directly affects the feature conditional probability, adjusting its contribution. The modified formula is as follows (taking the determination of whether a sample is an OSA patient as an example):

[0091]

[0092] Representative population type This indicates that, given an OSA-positive result and belonging to the G population, the characteristics... The conditional probability of occurrence; Representation of features Weighting coefficients; This represents the product of (weights × conditional probabilities) of all features. The higher the weight of a feature, the greater its contribution to the final molecule, thus making the posterior probability more biased towards the true class.

[0093] For example, suppose an overweight patient (BMI=32) drank alcohol at dinner that day, and the feature vector at the time of detection... Call the overweight patient parameter library to obtain... , Determine the weighting coefficients: The inherent weight of the population is ω1=1.4. The dynamic weighting of alcohol consumption is ω2=1.3.

[0094] Calculation of molecules (core part):

[0095]

[0096] If no weights are applied, the numerator is only... It is evident that after the weight adjustment, the molecular values ​​of the more critical features for overweight drinkers were significantly increased, making the posterior probability more reflective of the true OSA risk.

[0097] In this embodiment, the OSA prior probability of the corresponding crowd type of the to-be-detected person is first used as a basis, and then the extracted multi-dimensional features are matched with the corresponding feature conditional probability distribution, the high correlation features are given higher calculation weights through the calibration weights, and the posterior probabilities of the two results of suffering from OSA and not suffering from OSA are finally calculated and integrated, thereby providing a quantitative basis for subsequent classification, so as to finally realize the judgment of whether the to-be-detected person suffers from OSA.

[0098] In some embodiments, in step S102, the quality of each pressure signal is evaluated, and the target pressure signal is determined based on the quality evaluation result, including: calculating the energy and standard deviation of the sub-signal of each signal in the target frequency band to obtain the quality evaluation result in the target frequency band; determining the fusion weight of each signal according to the frequency band quality evaluation result, and weighting and fusing the same frequency band sub-signal of each signal based on the fusion weight of each signal to obtain the target pressure signal.

[0099] The target frequency band needs to be determined in combination with the sleep monitoring requirements, for example, it can be the respiratory main frequency band (0.1-0.5Hz) and the cardiovascular main frequency band (1-5Hz). For the sub-signal of each signal in each target frequency band, the energy and standard deviation are calculated, the energy reflects the signal strength of the sub-signal in the frequency band, and the higher the energy, the clearer the physiological signal in the frequency band, for example, the higher the energy of the respiratory frequency band, the more obvious the pressure fluctuation caused by respiration; the standard deviation reflects the stability of the sub-signal in the frequency band, and the smaller the standard deviation, the less the signal is disturbed. Through these two indicators, the frequency band quality evaluation result of each signal in each target frequency band can be obtained.

[0100] Then, the fusion weight of each signal is determined according to the frequency band quality evaluation result, and the core is to make the frequency band with better quality occupy a higher proportion in the fusion. For example, in the respiratory frequency band, if the quality evaluation score of A is 85 points and the quality evaluation score of B is 60 points, the fusion weight of A in the frequency band is set to 0.6 (85 / (85+60)), and the fusion weight of B is set to 0.4; in the heartbeat frequency band, the score of B is 80 points, and the score of A is 65 points, then the weight of B is set to 0.55, and the weight of A is set to 0.45, to ensure that the dominant frequency band of each signal can fully play a role. Finally, after determining the fusion weight corresponding to each signal, the sub-signal of each signal in the frequency band can be multiplied by the corresponding fusion weight and summed to obtain the fusion sub-signal of the frequency band, and then the fusion sub-signals of all target frequency bands are integrated to finally form a target pressure signal covering key physiological information, with high signal-to-noise ratio and strong stability, as the final target pressure signal.

[0101] In some embodiments, in step S102, the quality of each pressure signal is evaluated respectively, and the target pressure signal is determined based on the quality evaluation results, including the following steps.

[0102] It can be understood that the sliding window analysis of each signal for a preset time is to continuously capture the timing changes of the signal, and the preset time can be set to 20-40 seconds (for example, 30 seconds), and the sliding step can be set to 2-4 seconds. After each pressure signal is divided into a plurality of continuous and overlapping windows, the signal energy and standard deviation of each window can be calculated. Then, the energy fluctuation coefficient of each signal is determined based on the calculation formula of the energy fluctuation coefficient, and the calculation formula of the energy fluctuation coefficient can be: energy fluctuation coefficient = maximum difference of window energy / average value of window energy. Finally, based on the preset quality standard, the energy fluctuation coefficient and the standard deviation of each signal are determined whether the quality requirement is met, and the signal meeting the quality standard is preferentially selected. If multiple signals meet the quality requirement, the signal with smaller energy fluctuation coefficient is selected from the signals meeting the quality standard.

[0103] For example, two chest signals are analyzed by a 30-second sliding window, and the signal energy in the window is calculated.

[0104]

[0105] The standard deviation of the signal in the window is calculated.

[0106]

[0107] When the energy fluctuation coefficient of one signal is less than 0.2 and the standard deviation of the signal is less than 0.2, it is determined as a high-quality signal. If both satisfy, the signal with smaller energy fluctuation coefficient is selected; if only one satisfies, the signal is selected; if neither satisfies, a warning is issued and the cache segment with the best quality in the previous 30 seconds is enabled.

[0108] In some embodiments, as shown in FIG. 3, Figure 3 In step S103, the target pressure signal is preprocessed, and the features in time domain and frequency domain of the preprocessed target pressure signal are extracted, including the following steps S301-S304.

[0109] ​S301: Remove low-frequency noise below a predetermined frequency by adaptive Kalman filtering, and obtain a plurality of frequency band signals by multi-layer wavelet decomposition, wherein the plurality of frequency band signals include a respiratory main frequency band signal and a cardiovascular main frequency band signal.

[0110] In the application, low-frequency noise below 0.1 Hz, such as bed body vibration, can be removed by adaptive Kalman filtering. Then, 5 frequency band signals (0.1-0.5 Hz, 0.5-1 Hz, 1-2 Hz, 2-5 Hz, 5-10 Hz) are obtained by 5-layer wavelet decomposition, wherein the 0.1-0.5 Hz frequency band is the respiratory main frequency band signal, and the 1-2 Hz frequency band is the cardiovascular main frequency band signal.

[0111] S302: Detect the respiratory main frequency band signal by a peak detection algorithm to identify a respiratory wave peak, determine a respiratory threshold based on the signal mean value of the respiratory main frequency band signal, determine the respiratory frequency and the respiratory cycle variation coefficient according to the identified respiratory wave peak and the respiratory threshold, and extract the maximum amplitude, the minimum amplitude and the amplitude ratio of the respiratory signal from the respiratory main frequency band signal.

[0112] In the application, the respiratory main frequency band signal can be smoothed to weaken residual noise, and then a local maximum value judgment condition is set, that is, a certain data point is greater than each of the 3-5 adjacent data points before and after it. At the same time, combined with the overall strength of the signal, false peaks are excluded, and finally the highest point of pressure fluctuation in each respiratory cycle, that is, the respiratory wave peak, is located. Then, the respiratory threshold is determined based on the signal mean value of the respiratory main frequency band signal. The mean value of the signal in the entire analysis period can be calculated first, which can reflect the baseline level of respiratory pressure fluctuation. Then, the respiratory threshold is set based on the signal mean value, for example, it can be set to 1.5 times the signal mean value, to avoid wave peak misjudgment or omission caused by baseline drift.

[0113] Subsequently, the respiratory frequency and the respiratory cycle variation coefficient are determined according to the identified respiratory wave peak and the threshold. The respiratory frequency calculation needs to first count the time interval of adjacent two effective respiratory wave peaks, and then convert the number of respiratory cycles in 1 minute to the number of breaths per minute. The respiratory cycle variation coefficient is obtained by calculating the standard deviation of all respiratory cycles, and then dividing the cycle mean value. This coefficient can reflect the stability of respiratory rhythm. Finally, the maximum amplitude, the minimum amplitude and the amplitude ratio of the respiratory signal are extracted from the respiratory main frequency band signal. The amplitude is the pressure difference between the peak and the trough in a single respiratory cycle, the maximum amplitude is the maximum value of all amplitudes in the analysis period, the minimum amplitude is the minimum value, and the amplitude ratio is the ratio of the maximum amplitude to the minimum amplitude. These indicators can reflect the fluctuation of respiratory intensity. It should be noted that the amplitude ratio is usually <0.3 during OSA attack.

[0114] The calculation formula of the respiratory cycle variation coefficient can be as follows:

[0115] CV R =std(TR mean(T R )

[0116] wherein, T R is the respiratory cycle, std(T R ) is the standard deviation of the respiratory cycle, mean(T R ) represents the cycle mean.

[0117] S303: wavelet denoising is performed on the cardiovascular main frequency band signal to extract the heartbeat signal, and a heart rate variability index is calculated according to the heartbeat signal and a preset time window.

[0118] wherein, in the cardiovascular feature extraction link of obstructive sleep apnea (OSA) detection, first, wavelet denoising is performed on the cardiovascular main frequency band signal obtained through the multi-layer wavelet decomposition. The signal is first decomposed into high-frequency detail coefficients and low-frequency approximation coefficients of different scales. The high-frequency detail coefficients are processed by soft thresholding, that is, the coefficients less than the threshold value are set to zero, and the coefficients greater than the threshold value are contracted to retain the effective coefficients related to the heartbeat. Then, the signal is recovered through wavelet reconstruction, and finally the periodic heartbeat signal is extracted from the denoised signal. Then, the heart rate variability (HRV) index is calculated in combination with a preset time window. The preset time window needs to consider the timeliness and stability of HRV analysis. In sleep monitoring, it is usually set to 5-10 minutes (for example, 5 minutes). First, the time interval between two adjacent heartbeat peaks, i.e. RR interval, is identified from the heartbeat signal, and all RR interval data in the preset time window are counted. Then, based on these data, time domain indexes such as SDNN are calculated. SDNN is the standard deviation of RR interval, reflecting the overall variability of heart rate. OSA patients often have reduced SDNN due to autonomic nervous dysfunction.

[0119] S304: Perform fast Fourier transform on the selected period signal from the respiratory main frequency band signal and the cardiovascular main frequency band signal, and calculate the energy proportion of the respiratory main frequency band and the energy proportion of the cardiovascular main frequency band, respectively.

[0120] In the application, a continuous signal segment without obvious interference within 10-40 seconds can be selected, for example, a 30-second segment signal can be selected. After selecting the time period signal, fast Fourier transform (FFT) is performed on the selected time period signal. Fast Fourier transform can convert the pressure signal in the time domain into the frequency-energy spectrum in the frequency domain, which intuitively presents the signal energy distribution corresponding to different frequency components, wherein the abscissa is the frequency and the ordinate is the energy value. Through FFT, the energy contribution of each subdivided frequency in the respiratory and cardiovascular main frequency band can be accurately captured. For the respiratory main frequency band, first, the frequency interval of 0.1-0.5 Hz is located in the frequency domain spectrum, the total energy of all frequency points in the interval is calculated, the total energy of the full frequency band of the entire frequency domain spectrum is calculated, and finally the energy ratio of the respiratory main frequency band is obtained by “respiratory main frequency band total energy ÷ full frequency band total energy × 100%”. The calculation logic of the cardiovascular main frequency band energy ratio is consistent, only the frequency interval is replaced by 1-2 Hz, and the ratio of the total energy of the interval to the total energy of the full frequency band is calculated.

[0121] Exemplarily, for 30-second signal FFT transformation (point number 8192), the energy ratio of the respiratory main frequency band 0.1-0.5 Hz is calculated according to the following formula:

[0122] P R =E 0.1-0.5HZ / E total

[0123] Wherein, P R represents the energy ratio of the respiratory main frequency band, E 0.1-0.5HZ represents the total energy of the respiratory main frequency band, and E total represents the total energy of the full frequency band. The calculation method of the cardiovascular main frequency band energy ratio PH is similar. Wherein, the decrease of PR under OSA state is > 30%, and the fluctuation amplitude of PH is increased by > 20%.

[0124] Based on the same inventive concept, as shown in Figure 4 , the application also provides a obstructive sleep apnea detection system 400, comprising: a processing module 402 and at least two bed-mounted pressure sensors 401.

[0125] Each bed-mounted pressure sensor 401 is adapted to be installed at a position corresponding to the chest of a human body on a mattress. The processing module 402 is connected with each bed-mounted pressure sensor 401, and the processing module 402 is used to execute the obstructive sleep apnea detection method according to any one of the above schemes.

[0126] The obstructive sleep apnea detection system 400 described above can detect at least two pressure signals of the body in a sleeping state through the bed-mounted pressure sensors 401 installed on the mattress corresponding to the chest of the human body. The processing module 402 can perform quality assessment on each pressure signal after receiving each pressure signal, determine a target pressure signal based on the quality assessment result, and obtain a target pressure signal with better quality. The processing module 402 can pre-process the target pressure signal after obtaining the target pressure signal, and then extract features in the time domain and the frequency domain from the pre-processed target pressure signal to accurately extract multi-dimensional features related to OSA from the pressure signal. On this basis, the processing module 402 inputs the extracted multi-dimensional features, the population type characteristics and the pre-sleep activity characteristics of the to-be-detected person into the pre-trained machine learning model, which can improve the diagnostic accuracy through multi-dimensional physiological characteristics. At the same time, the model can break through the limitation of a single population standard by relying on the population type characteristics, optimize the judgment logic for high-risk populations and pre-sleep activities, enhance the individual adaptation ability, and further improve the classification accuracy of OSA. Moreover, the obstructive sleep apnea detection system 400 of the present application does not need to rely on complex and expensive PSG equipment, and at least two bed-mounted pressure sensors 401 are arranged on the mattress, so that the signal can be collected through the bed-mounted pressure sensor 401. It is suitable for primary medical care and home screening scenarios, and has high convenience.

[0127] In some embodiments, the processing module 402 includes a signal conditioning module, a low-pass filter, a high-pass filter, a data acquisition module, and a data processing module.

[0128] The signal conditioning module is used to receive the pressure signals output by the bed-mounted pressure sensors 401 and amplify each pressure signal to obtain an amplified pressure signal. The low-pass filter is used to filter out high-frequency noise of the amplified pressure signal. The high-pass filter is used to filter the signal output by the low-pass filter to suppress baseline drift. The data acquisition module is used to perform analog-to-digital conversion on the signal output by the high-pass filter to generate a digital signal. The data processing module is used to process the digital signal output by the data acquisition module and perform the obstructive sleep apnea detection method according to any of the above solutions.

[0129] The signal conditioning module uses a gain-adjustable instrument amplifier to amplify the two signals to 2-5V. The low-pass filter can be a 5th order Butterworth low-pass filter. The data acquisition module can complete A / D conversion at a sampling rate of 500Hz to generate a 16-bit digital signal.

[0130] It can be understood that the bed-mounted pressure sensor 401 sensor is limited by the detection principle, and the directly output pressure signal is usually in the millivolt level, which is easy to be covered by environmental noise in the transmission process, therefore, it is necessary to accurately amplify each original signal through the instrument amplifier built-in the signal conditioning module, while suppressing the noise interference of the amplifier itself, to ensure that the amplified pressure signal can clearly retain the characteristics of the small pressure fluctuations caused by breathing and heartbeat, and lay a foundation for subsequent filtering. Then, the low-pass filter filters out the "high-frequency noise" mixed in the amplified pressure signal, and the cutoff frequency can be set to 30Hz, only allowing useful physiological signals below the frequency to pass through, and blocking the interference of high-frequency noise on the signal baseline. Subsequently, the high-pass filter focuses on solving the "baseline drift" problem. The low-frequency drift of the low-pass filtered signal may be caused by slow changes in sensor temperature, long-term slight deformation of the bed body, or slow turning of the human body, and the module filters out such low-frequency drift components by setting a cutoff frequency of about 0.05Hz, so that the signal baseline returns to stable, and ensures that the extracted breathing and heartbeat features are not affected by the baseline fluctuation. Subsequently, the data acquisition module samples and quantizes the analog signal after high-pass filtering through the built-in analog-to-digital converter (ADC), and finally generates a digital signal that can be transmitted and calculated. Finally, the data processing module, as the core link of signal processing, after receiving the digital signal output by the data acquisition module, will perform the key operations in the previous process, realizing the conversion from the original pressure signal to the sleep health assessment result.

[0131] It should be noted that the obstructive sleep apnea detection system 400 provided by the embodiments of the present application is based on the same application concept as the obstructive sleep apnea detection method provided by the embodiments of the present application, and therefore the specific implementation of this embodiment can be referred to the implementation of the aforementioned obstructive sleep apnea detection method, and the repeated parts will not be described again.

[0132] In some embodiments, an electronic device provided by the embodiments of the present application includes a processor and a memory; the memory has a computer program stored therein, wherein the computer program, when executed by the processor, implements the obstructive sleep apnea detection method described above.

[0133] Specifically, the processor may, for example, include a general-purpose microprocessor, an instruction set processor, and / or a related chipset, and / or a special-purpose microprocessor (such as an application-specific integrated circuit (ASIC)), etc. The processor can also include on-board memory for cache use. The processor can be a single processing unit for performing different actions of the method processes according to the embodiments of the present application or a plurality of processing units.

[0134] Memory, for example, can be any media capable of storing, containing, communicating, propagating, or transporting instructions. For example, memory can include, without limitation, electronic, magnetic, optical, electromagnetic, infrared, or semiconductor systems, apparatus, devices, or propagation mediums. Specific examples of memory include magnetic storage devices such as hard disks or hard disk drives (HDDs); optical storage devices such as compact disc (CD-ROM); and / or random access memory (RAM) or flash memory; and / or wired / wireless communication links.

[0135] The application also provides a computer readable medium having stored thereon a computer program which, when executed by a processor, implements the obstructive sleep apnea detection method described above. The computer readable medium can be included in the device / apparatus / system described in the above embodiments; or can exist separately and not be assembled into the device / apparatus / system. The computer readable medium carries one or more programs which, when executed, implement the method of the embodiments of the application.

[0136] According to an embodiment of the application, the computer readable medium can be a computer readable signal medium or a computer readable storage medium or any combination thereof. The computer readable storage medium can be, for example, but not limited to, an electronic, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus, or device, or any suitable combination thereof. More specific examples of the computer readable storage medium can include, but are not limited to, an electrical connection having one or more wires, a portable computer diskette, a hard disk, a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or flash memory), an optical fiber, a portable compact disc read-only memory (CD-ROM), an optical storage device, a magnetic storage device, or any suitable combination thereof. In this application, the computer readable storage medium can be any tangible medium that contains or stores a program used by or in connection with an instruction execution system, apparatus, or device. In this application, the computer readable signal medium can include a computer readable program code propagated in or on a carrier medium, which can be any medium used to transmit the computer readable program code for use by or in connection with an instruction execution system, apparatus, or device. The computer readable program code propagated in or on the carrier medium can take any form capable of propagating the computer readable program code, including but not limited to, an electromagnetic signal, an optical signal, or any suitable combination thereof. The computer readable medium can also be any computer readable medium other than the computer readable storage medium, which can transmit, propagate, or transport the program for use by or in connection with the instruction execution system, apparatus, or device. The program code contained in the computer readable medium can be transmitted in any suitable format, including but not limited to, wireless, wired, optical, radio frequency (RF), or any suitable combination thereof.

[0137] It is understood by those skilled in the art that features recited in the various embodiments of the present application can be combined and / or incorporated in various combinations, even if such combinations have not been explicitly recited in the present application. In particular, features recited in the various embodiments of the present application can be combined and / or incorporated in various combinations without departing from the spirit and teachings of the present application. All such combinations and / or incorporations are within the scope of the present application. Therefore, the scope of the present application should not be limited to the above-described embodiments. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present application should be included in the scope of the present application.

Claims

1. A method for processing obstructive sleep apnea signals, characterized in that, include: Acquire pressure signals detected by at least two bed-mounted pressure sensors during human sleep, as well as the population type characteristics and pre-sleep activity characteristics of the person being tested; Each pressure signal is evaluated for quality, and the target pressure signal is determined based on the evaluation results. The target pressure signal is preprocessed, and time-domain and frequency-domain features are extracted from the preprocessed target pressure signal. The extracted multi-dimensional features, the population type features of the person to be tested, and the pre-sleep activity features are input into the Bayesian model, so that the Bayesian model can call the population-specific feature parameter library according to the population type features of the person to be tested. Obtain the population type weight coefficient and activity weight coefficient corresponding to the population type characteristics and pre-sleep activity characteristics of the person to be tested, respectively; The calibration weights for each time-domain feature and each frequency-domain feature are determined based on the population type weight coefficient and the activity weight coefficient. Based on the population type of the person to be tested, retrieve the characteristic conditional probability distribution of the corresponding population from the corresponding population-specific characteristic parameter database; The posterior probabilities of whether the person under test has obstructive sleep apnea are determined based on the calibration weights and feature conditional probability distributions of the extracted multi-dimensional features, time-domain features, and frequency-domain features. The classification result of whether the person being tested has obstructive sleep apnea is determined based on the posterior probability of the two types of results.

2. The obstructive sleep apnea signal processing method as described in claim 1, characterized in that, The process of evaluating the quality of each pressure signal and determining the target pressure signal based on the evaluation results includes: The energy and standard deviation of each signal sub-signal in the target frequency band are calculated to obtain the quality assessment results in the target frequency band. The fusion weights of each signal are determined based on the frequency band quality assessment results, and the sub-signals of each signal in the same frequency band are weighted and fused based on the fusion weights of each signal to obtain the target pressure signal.

3. The obstructive sleep apnea signal processing method as described in claim 1, characterized in that, The process of evaluating the quality of each pressure signal and determining the target pressure signal based on the evaluation results includes: Perform a sliding window analysis for each signal with a preset time, and calculate the signal energy and standard deviation within the window; The corresponding energy fluctuation coefficient is determined based on the signal energy within the window of each signal. The optimal signal is determined based on the energy fluctuation coefficient and standard deviation of each signal, and the optimal signal is used as the target pressure signal.

4. The obstructive sleep apnea signal processing method as described in claim 1, characterized in that, The preprocessing of the target pressure signal, and the extraction of time-domain and frequency-domain features from the preprocessed target pressure signal, includes: Low-frequency noise below a predetermined frequency is removed by adaptive Kalman filtering, and multiple frequency band signals are obtained by multi-layer wavelet decomposition, wherein the multiple frequency band signals include respiratory main frequency band signals and cardiovascular main frequency band signals. The respiratory main frequency band signal is detected by a peak detection algorithm to identify respiratory peaks. A respiratory threshold is determined based on the signal mean of the respiratory main frequency band signal. The respiratory frequency and respiratory cycle variation coefficient are determined according to the identified respiratory peaks and the respiratory threshold. The maximum amplitude, minimum amplitude and amplitude ratio of the respiratory signal are extracted from the respiratory main frequency band signal. Wavelet denoising is performed on the cardiovascular main frequency band signal to extract the heartbeat signal, and the heart rate variability index is calculated based on the heartbeat signal and a preset time window; Fast Fourier Transform is performed on time period signals selected from the respiratory dominant frequency band signal and the cardiovascular dominant frequency band signal to calculate the energy proportion of the respiratory dominant frequency band and the energy proportion of the cardiovascular dominant frequency band, respectively.

5. A signal processing system for obstructive sleep apnea, characterized in that, include: Processing module and at least two bed-mounted pressure sensors; Each bed pressure sensor is designed to be installed on the mattress at the position corresponding to the human chest. The processing module is connected to each bed pressure sensor, and the processing module is used to execute the obstructive sleep apnea signal processing method as described in any one of claims 1 to 4.

6. The obstructive sleep apnea signal processing system as described in claim 5, characterized in that, The processing module includes a signal conditioning module, a low-pass filter, a high-pass filter, a data acquisition module, and a data processing module; The signal conditioning module is used to receive the pressure signals output by each bed pressure sensor and amplify each pressure signal to obtain an amplified pressure signal. The low-pass filter is used to filter out high-frequency noise in the amplified pressure signal; The high-pass filter is used to filter the signal output by the low-pass filter in order to suppress baseline drift; The data acquisition module is used to perform analog-to-digital conversion on the signal output by the high-pass filter to generate a digital signal; The data processing module is used for the digital signals output by the data acquisition module and executes the obstructive sleep apnea signal processing method as described in any one of claims 1 to 4.

7. An electronic device, characterized in that, It includes a processor and a memory; the memory stores a computer program, wherein the computer program, when executed by the processor, implements the obstructive sleep apnea signal processing method as described in any one of claims 1 to 4.

8. A computer storage medium, characterized in that, It stores a computer program thereon, wherein the computer program, when executed by a processor, implements the obstructive sleep apnea signal processing method as described in any one of claims 1 to 4.

Images