Snore detection system applied to oral appliance

Abstract

The invention discloses a snore detection system applied to an oral appliance, which comprises the oral appliance arranged in the oral cavity of a detector, a smart phone arranged outside the detector and wirelessly connected with the oral appliance, and a snore detection software program installed in the smart phone, and the cloud server is wirelessly connected with the smart phone. The snore recording and analyzing device has the effects of recording and analyzing snore and providing more accurate and reliable data support for the oral appliance.

Description

Technical Field

[0001] The invention belongs to the field of medical equipment, and in particular relates to a snoring detection system applied to an oral appliance. Background Art

[0002] With the advancement of science and technology and the increasing demand for health management, the future snoring detection and oral appliance effect evaluation will develop towards higher accuracy and systematization: 1. Integrated system: Future systems may integrate more advanced audio analysis technologies to provide comprehensive snoring analysis and effect evaluation. 2. Data-driven evaluation: Using big data and machine learning technology, in-depth analysis and prediction of snoring changes can be achieved, thereby improving the accuracy of oral appliance effect evaluation. 3. Personalized treatment: Combined with individualized snoring analysis data, customized treatment plans and real-time feedback are provided to further optimize the use effect of oral appliances. At present, the snoring detection and management devices on the market mainly rely on traditional sleep monitoring. These devices usually focus on monitoring apnea, sleep cycle and other related parameters. However, these devices have relatively limited detection and analysis functions for snoring and cannot provide a comprehensive evaluation of snoring changes. Now, more and more sound processing and analysis technologies are being applied to snoring monitoring. Some advanced sleep monitoring devices have integrated sound recording functions, which are used to detect the presence and frequency of snoring. However, the snoring monitoring technology lacks systematic analysis of snoring changes and quantitative evaluation of treatment effects. In addition, the final evaluation results of the monitoring of snoring prevention effects combined with the application of oral appliances mainly rely on the doctor's experience and the patient's subjective feedback. Therefore, this part of the technology also lacks systematic data support. Summary of the invention

[0003] In view of the above problems, an object of the present invention is to provide a snoring detection system for use in an oral appliance for recording and analyzing snoring and providing more accurate and reliable data support for the oral appliance.

[0004] To achieve the above-mentioned purpose, the present invention provides a snoring detection system applied to an oral appliance, which is characterized by comprising an oral appliance arranged in the mouth of a person being tested, a smart phone arranged outside the person being tested and wirelessly connected to the oral appliance, a snoring detection software program installed in the smart phone, and a cloud server wirelessly connected to the smart phone.

[0005] The snoring detection software program includes an audio acquisition module that is connected to the microphone inside the smartphone and is used to adopt audio data and connect to a cloud server to upload the audio data; a data preprocessing module is provided in the smartphone and is wirelessly connected to the cloud server and performs denoising and signal enhancement processing on the audio data after downloading; a feature extraction module is provided in the smartphone and is connected to the data preprocessing module and extracts the preprocessed audio signal for snoring identification and evaluation of key features; a snoring detection module is provided in the smartphone and is connected to the feature extraction module and accurately identifies snoring events in the audio data; an effect evaluation module is provided in the smartphone and is connected to the snoring detection module and to the oral appliance respectively and performs quantitative analysis and evaluation on the effect of the use by the detector; and a visual report generation module is provided in the smartphone and is connected to the effect evaluation module.

[0006] The visual report generation module presents the effect evaluation results in the form of charts on the display screen of the smartphone.

[0007] In some embodiments, the data preprocessing module includes a bandpass filter for removing high noise, and an adaptive filter connected to the bandpass filter.

[0008] The bandpass filter removes noises above 8000 Hz and below 20 Hz, because snoring sounds are mainly concentrated between 20 Hz and 2000 Hz. The adaptive filter further removes environmental noise from the filtered signal.

[0009] In some embodiments, the implementation of the adaptive filter is based on the LMS algorithm, which is used to estimate the parameters of the system to minimize the mean square error between the output signal and the desired signal, and its formula is:

[0010] w(n+1)=w(n)+2μe(n)x(n)

[0011] Where: w(n) is the weight vector of the filter, μ is the step size factor used to control the rate of weight update, e(n) is the current error signal, and x(n) is the input signal;

[0012] By iteratively updating the weights, the adaptive filter can adaptively remove noise and enhance the snoring signal.

[0013] In some embodiments, the feature extraction module uses Mel-frequency cepstral coefficients (MFCCs).

[0014] The calculation steps of Mel frequency cepstral coefficients MFCCs are as follows:

[0015] (1) The extraction process of MFCCs includes the following key steps:

[0016] A. Pre-emphasis:

[0017] First, the input audio signal is pre-emphasized to enhance the high-frequency components. After that, the high-frequency attenuation caused by the damping effect of the vocal organs in the speech signal is offset. The pre-emphasis filter is a first-order high-pass filter, and its mathematical expression is:

[0018] y(t)=x(t)-αx(t-1)

[0019] Where: y(t) is the pre-emphasized signal, x(t) is the original input signal, and α is the pre-emphasis coefficient;

[0020] B. Framing and windowing:

[0021] The audio signal is non-stationary, and the framing method divides it into multiple short-term stable frames with 50% overlap between frames. On the basis of framing, a Hamming window is applied to each frame signal to reduce spectrum leakage. The expression is:

[0022]

[0023] Where: N is the length of the window, n is the current time point.

[0024] C. Fast Fourier Transform:

[0025] Perform fast Fourier transform on each frame of the windowed signal to convert it from the time domain to the frequency domain; the result of FFT is the power spectrum of the signal, which is expressed as:

[0026]

[0027] Where X(k) is the kth frequency component, x(n) is the input signal, and N is the number of FFT points.

[0028] D. Mel frequency cepstrum coefficients: Use the Mel filter bank to weight the FFT result to generate the Mel spectrum. The Mel scale is a nonlinear frequency scale that conforms to human auditory perception and is defined as:

[0029]

[0030] Where: m is the Mel frequency, f is the linear frequency (Hz).

[0031] After converting the linear frequency to the Mel frequency, a set of triangular filters are applied on this scale to perform weighted summation on the power spectrum of the FFT to obtain the Mel spectrum coefficients. Then, through logarithmic operations and discrete cosine transform (DCT), the Mel frequency cepstral coefficients (MFCCs) are obtained. DCT converts the features of the Mel spectrum from the frequency domain to the cepstral domain and compresses it into a set of low-dimensional feature vectors. MFCC is calculated as follows:

[0032]

[0033] Where: m represents the index of the cepstral coefficient, and K is the number of Mel filter banks.

[0034] (2) Feature optimization and dimensionality reduction

[0035] In order to improve the distinguishing ability and computational efficiency of features, the principal component analysis (PCA) dimensionality reduction technique is also used to optimize the extracted MFCCs features. PCA projects high-dimensional feature vectors into low-dimensional space through linear transformation to retain the main information of the data. The optimization process of PCA is as follows:

[0036] A. Calculate the covariance matrix: Evaluate the correlation between features by calculating the covariance matrix of MFCCs features.

[0037]

[0038] Where: C is the covariance matrix, N is the number of samples, x i is the eigenvector of the i-th sample, and μ is the mean of the eigenvector.

[0039] B. Eigenvalue decomposition: Perform eigenvalue decomposition on the covariance matrix to obtain eigenvalues ​​and corresponding eigenvectors. Select the eigenvector corresponding to the maximum eigenvalue as the new feature space after dimensionality reduction.

[0040] Cv i =λ i v i

[0041] Where: i is the i-th eigenvalue, v i is the corresponding feature vector.

[0042] Finally, the MFCCs feature vector is projected into a new low-dimensional feature space through linear transformation to obtain the optimized feature vector.

[0043] In some embodiments, the snoring detection module adopts a model based on a convolutional neural network, which can effectively capture the timing and frequency characteristics of the audio signal.

[0044] In some embodiments, the model is based on a convolutional neural network whose input is the extracted MFCCs features and corresponding labels, and after passing through multiple convolutional layers and pooling layers, a high-dimensional feature map is extracted. The process is as follows:

[0045] (1) The extracted MFCCs features are firstly subjected to spatial feature extraction through multiple convolutional layers; assuming that the output of the lth convolutional layer is H (l) , then it is calculated by the following formula:

[0046] H (l) =σ(W (l) *H (l-1) +b (l) )

[0047] Among them, W (l) is the lth layer of convolution, * indicates the convolution operation, b (l) is the bias term, and σ is the activation function (ReLU).

[0048] (2) After each convolutional layer, a pooling layer is applied to reduce the dimension of the feature map and extract the main features. The commonly used pooling operation is maximum pooling, and its calculation formula is:

[0049]

[0050] Among them, R is the pooling window area.

[0051] After multiple layers of convolution and pooling, a high-dimensional feature map is obtained; the feature map is flattened and input into the fully connected layer, and finally a sigmoid activation function is used to output the probability value of each time segment containing a snoring event.

[0052]

[0053] Among them, h is the flattened eigenvector, W FC and b FC are the weight and bias of the fully connected layer respectively.

[0054] The network loss function uses the binary cross entropy loss function:

[0055]

[0056] Where: N is the total number of samples, y i is the true label (0 or 1) of the i-th sample, p i is the probability of snoring predicted by the model.

[0057] The model is trained through back-propagation and stochastic gradient descent algorithms to minimize the loss function and optimize the model parameters.

[0058] In some implementations, the snoring detection module is also provided with a model optimization function.

[0059] The model optimization function allows users to evaluate the detection results through the feedback interface provided by the system, generate a feedback data set, combine the feedback data with the original training data to form an extended training set, use the extended training set to retrain the model, and further adjust the model parameters through the optimization algorithm Adam optimizer to adapt to the new data distribution and user needs; regularly evaluate the performance of the optimized model to ensure the detection effect of the model in different environments and conditions.

[0060] In some embodiments, the effect evaluation module includes AHI evaluation and snoring characteristic change evaluation.

[0061] AHI assessment is an important indicator for measuring the severity of sleep apnea, which is defined as the sum of the number of apneas and hypopneas occurring per hour of sleep.

[0062] The specific calculation is disclosed as follows:

[0063]

[0064] The effect evaluation module detects apnea and hypopnea events in the signal processing and feature extraction modules. The start and end time of each event is marked by analyzing the energy changes, breathing patterns and audio feature MFCCs of the audio signal. The AHI value is calculated based on the number of apnea and hypopnea events detected and the patient's total sleep time. In order to improve the accuracy of the calculation, the AHI is calculated within a specific time window (every hour), and then the AHI over the entire sleep cycle is averaged. The therapeutic effect of the appliance is evaluated by comparing the AHI values ​​before and after the use of the appliance. If the AHI value is significantly reduced after using the appliance, especially if the AHI drops to the normal range during sleep (AHI <5), it indicates that the appliance is effective in improving sleep breathing disorders.

[0065] The snoring characteristics change assessment is to further verify the effect of the appliance by evaluating the changes in snoring characteristics. The snoring characteristics include snoring frequency, snoring duration, and snoring intensity. By comparing the changes in these indicators before and after the use of the appliance, combined with the evaluation of the AHI value, the system can provide a comprehensive evaluation result to judge the clinical effect of the appliance; ultimately, these evaluation data will be presented in the form of charts and reports to help doctors and patients intuitively understand the treatment effect and provide a basis for further treatment plans.

[0066] In some embodiments, the visual report generation module is an AHI trend graph or a snore spectrum graph or a calendar view or a statistical chart.

[0067] The AHI trend chart shows the changing trend of the patient's AHI value before and after the use of the orthodontic appliance.

[0068] Snoring spectrum: displays the frequency distribution of snoring and its changing trend.

[0069] Calendar view: displays the frequency and intensity of snoring each day;

[0070] The charts include bar charts and line charts, which are used to show the changes in snoring assessment indicators in different time periods.

[0071] The beneficial effects of the present invention are that it can be used to record and analyze snoring and provide more accurate and reliable data support for oral appliances. Since the snoring detection software program realizes the collection, processing, analysis and effect evaluation of audio data through a smart phone software program, it provides more accurate and reliable data support for the effect evaluation of oral appliances, promoting the development and application of this field. That is, it quantifies the evaluation process, provides an intuitive and easy-to-use operation interface, improves the user experience, and enables patients and medical staff to conduct effect evaluation more conveniently. Through more accurate evaluation results, it can help doctors better adjust the treatment plan, thereby improving the treatment effect of oral appliances and improving the quality of life of patients. The specific effects are: (1) Objective and accurate effect evaluation: Through advanced audio recording and analysis technology, the changes in snoring can be accurately recorded and analyzed. Unlike traditional methods that rely on subjective feedback, the snoring detection software program provides an objective evaluation based on data and can quantify the treatment effect of oral appliances. (2) Systematic snoring analysis: The snoring detection software program can perform a detailed analysis of snoring, including multiple parameters such as AHI, volume, frequency, time, etc. This systematic analysis method is more comprehensive than the existing simplified detection technology and helps to fully understand the changes in snoring. (3) Personalized treatment feedback: By comparing the snoring data before and after using the oral appliance, personalized treatment feedback is provided. This enables doctors to adjust the treatment plan according to the specific changes in snoring and improve the treatment effect. (4) Efficient effect tracking: Patients can monitor and track the effects of the oral appliance in real time during the use of the oral appliance, and promptly identify problems in the treatment and make adjustments. This not only improves the continuity and effectiveness of the treatment, but also reduces possible side effects or discomfort. (5) Data-driven optimization suggestions: Based on the collected snoring data, detailed reports and optimization suggestions are generated to help doctors and patients develop more effective treatment plans. Data-driven optimization suggestions help personalized treatment strategies and decisions. (6) Simplified user experience: The snoring detection software program is easy to use, and patients can easily record snoring and obtain evaluation results. The simplified user experience improves patient compliance and treatment effects. This achieves the effect of recording and analyzing snoring and providing more accurate and reliable data support for oral appliances. BRIEF DESCRIPTION OF THE DRAWINGS

[0072] Figure 1 is a flow chart of the present invention;

[0073] Figure 2 This is a curve chart of the data of subject 1 before and after using the appliance for 7 days;

[0074] Figure 3 is a curve chart of the sleep data of the tester 1;

[0075] Figure 4 This is a graph showing the data of subject 2 before and after using the appliance for 7 days;

[0076] Figure 5 This is a curve chart of the sleep data of subject 2. DETAILED DESCRIPTION

[0077] The invention will be further described in detail below in conjunction with the accompanying drawings.

[0078] like Figure 1 As shown, a snoring detection system applied to an oral appliance includes an oral appliance installed in the oral cavity of a tester, a smart phone wirelessly connected to the oral appliance installed outside the tester, a snoring detection software program installed in the smart phone, and a cloud server wirelessly connected to the smart phone. The snoring detection software program mainly realizes the collection, processing, analysis and effect evaluation of audio data through the smart phone software program.

[0079] The snoring detection software program includes an audio acquisition module connected to a microphone inside a smartphone and used to adopt audio data and connect to a cloud server to upload the audio data, a data preprocessing module wirelessly connected to a cloud server in the smartphone and denoising and signal enhancement processing after the audio data is downloaded, a feature extraction module connected to the data preprocessing module in the smartphone and extracting the preprocessed audio signal for snoring recognition and evaluation of key features, a snoring detection module connected to the feature extraction module in the smartphone and accurately identifying snoring events in the audio data, an effect evaluation module connected to the snoring detection module and to the oral appliance in the smartphone and quantitatively analyzing and evaluating the effect of the tester, and a visual report generation module connected to the effect evaluation module in the smartphone. The audio acquisition module collects the audio data of the patient during sleep in real time through the microphone of the smartphone. The audio data is transmitted to the cloud server through a secure wireless transmission protocol for further processing. The function of the data acquisition unit is limited to audio recording and data uploading, and does not involve data preprocessing or analysis tasks, thereby ensuring the originality and integrity of the audio data.

[0080] The data preprocessing module includes a bandpass filter for removing high noise and an adaptive filter connected to the bandpass filter. The bandpass filter removes noise above 8000Hz and below 20Hz, because snoring is mainly concentrated between 20Hz and 2000Hz; the adaptive filter further removes environmental noise from the filtered signal, such as continuous interference sounds such as air conditioners and fans.

[0081] The implementation of the adaptive filter is based on the LMS (Least Mean Squares) algorithm, which is used to estimate the parameters of the system to minimize the mean square error between the output signal and the expected signal. The formula is:

[0082] w(n+1)=w(n)+2μe(n)x(n)

[0083] Where: w(n) is the weight vector of the filter, μ is the step factor used to control the rate of weight update, e(n) is the current error signal, and x(n) is the input signal.

[0084] By iteratively updating the weights, the adaptive filter can adaptively remove noise and enhance the snoring signal.

[0085] The feature extraction module uses Mel frequency cepstral coefficients MFCCs. The calculation steps of Mel frequency cepstral coefficients MFCCs are as follows:

[0086] (1) The extraction process of MFCCs includes the following key steps:

[0087] A. Pre-emphasis:

[0088] First, the input audio signal is pre-emphasized to enhance the high-frequency components; then, the high-frequency attenuation caused by the damping effect of the vocal organs in the speech signal is offset; the pre-emphasis filter is a first-order high-pass filter, and its mathematical expression is:

[0089] y(t)=x(t)-αx(t-1)

[0090] Where: y(t) is the pre-emphasized signal, x(t) is the original input signal, and α is the pre-emphasis coefficient.

[0091] B. Framing and windowing:

[0092] The audio signal is non-stationary. Framing divides it into multiple short-term stable frames with 50% overlap between frames. On the basis of framing, a Hamming window is applied to each frame signal to reduce spectrum leakage. The expression is:

[0093]

[0094] Where: N is the length of the window, n is the current time point.

[0095] C. Fast Fourier Transform:

[0096] Perform fast Fourier transform on each frame of the windowed signal to convert it from the time domain to the frequency domain; the result of FFT is the power spectrum of the signal, which is expressed as:

[0097]

[0098] Where X(k) is the kth frequency component, x(n) is the input signal, and N is the number of FFT points.

[0099] D. Mel frequency cepstrum coefficients: Use the Mel filter bank to weight the FFT result to generate the Mel spectrum; the Mel scale is a nonlinear frequency scale that conforms to the human auditory perception and is defined as:

[0100]

[0101] Where: m is the Mel frequency, f is the linear frequency (Hz).

[0102] After converting the linear frequency to the Mel frequency, a set of triangular filters are applied on this scale to perform weighted summation on the power spectrum of the FFT to obtain the Mel spectrum coefficients; then, through logarithmic operations and discrete cosine transform (DCT), the Mel frequency cepstral coefficients (MFCCs) are obtained; DCT converts the features of the Mel spectrum from the frequency domain to the cepstral domain and compresses it into a set of low-dimensional feature vectors; MFCC is calculated as follows:

[0103]

[0104] Where: m represents the index of the cepstral coefficient, and K is the number of Mel filter banks.

[0105] (2) Feature optimization and dimensionality reduction

[0106] In order to improve the distinguishing ability and computational efficiency of features, the principal component analysis (PCA) dimensionality reduction technique is also used to optimize the extracted MFCCs features; PCA projects high-dimensional feature vectors into low-dimensional space through linear transformation to retain the main information of the data; the optimization process of PCA is as follows:

[0107] A. Calculate the covariance matrix: Evaluate the correlation between features by calculating the covariance matrix of MFCCs features;

[0108]

[0109] Where: C is the covariance matrix, N is the number of samples, x iis the eigenvector of the i-th sample, and μ is the mean of the eigenvector.

[0110] B. Eigenvalue decomposition: Perform eigenvalue decomposition on the covariance matrix to obtain eigenvalues ​​and corresponding eigenvectors; select the eigenvector corresponding to the maximum eigenvalue as the new feature space after dimensionality reduction;

[0111] Cv i =λ i v i

[0112] Where: i is the i-th eigenvalue, v i is the corresponding feature vector.

[0113] Finally, the MFCCs feature vector is projected into a new low-dimensional feature space through linear transformation to obtain the optimized feature vector.

[0114] The snoring detection module adopts a model based on a convolutional neural network, which can effectively capture the timing and frequency characteristics of audio signals.

[0115] The model is based on the convolutional neural network. The input is the extracted MFCCs features and corresponding labels. After multiple convolutional layers and pooling layers, high-dimensional feature maps are extracted. The process is as follows:

[0116] (1) The extracted MFCCs features are firstly subjected to spatial feature extraction through multiple convolutional layers; assuming that the output of the lth convolutional layer is H (l) , then it is calculated by the following formula:

[0117] H (l) =σ(W (l) *H (l-1) +b (l) )

[0118] Among them, W (l) is the lth layer of convolution, * indicates the convolution operation, b (l) is the bias term, and σ is the activation function (ReLU).

[0119] (2) After each convolutional layer, a pooling layer is applied to reduce the dimension of the feature map and extract the main features; the commonly used pooling operation is maximum pooling, and its calculation formula is:

[0120]

[0121] Among them, R is the pooling window area.

[0122] After multiple layers of convolution and pooling, a high-dimensional feature map is obtained. The feature map is flattened and input into the fully connected layer, and finally a sigmoid activation function is used to output the probability value of each time segment containing a snoring event.

[0123]

[0124] Among them, h is the flattened eigenvector, W FC and b FC are the weight and bias of the fully connected layer respectively.

[0125] The network loss function uses the binary cross entropy loss function:

[0126]

[0127] Where: N is the total number of samples, y i is the true label (0 or 1) of the i-th sample, p i is the probability of snoring predicted by the model.

[0128] The model is trained through back-propagation and stochastic gradient descent algorithms to minimize the loss function and optimize the model parameters.

[0129] The snoring detection module is also equipped with a model optimization function;

[0130] The model optimization function allows users to evaluate the detection results through the feedback interface provided by the system, generate a feedback data set, combine the feedback data with the original training data to form an extended training set, use the extended training set to retrain the model, and further adjust the model parameters through the optimization algorithm Adam optimizer to adapt to the new data distribution and user needs; regularly evaluate the performance of the optimized model to ensure the detection effect of the model in different environments and conditions.

[0131] The effect evaluation module includes AHI assessment and snoring characteristic change assessment;

[0132] AHI assessment is an important indicator for measuring the severity of sleep apnea, which is defined as the sum of the number of apneas and hypopneas occurring per hour of sleep.

[0133] The specific calculation is disclosed as follows:

[0134]

[0135] The effect evaluation module detects apnea and hypopnea events in the signal processing and feature extraction modules; the start and end time of each event is marked by analyzing the energy changes, breathing patterns and audio feature MFCCs of the audio signal; the AHI value is calculated based on the number of detected apnea and hypopnea events and the patient's total sleep time; in order to improve the accuracy of the calculation, the AHI is calculated within a specific time window (every hour), and then the AHI over the entire sleep cycle is averaged; the therapeutic effect of the appliance is evaluated by comparing the AHI values ​​before and after the use of the appliance; if the AHI value is significantly reduced after using the appliance, especially when the AHI drops to the normal range during sleep (AHI <5), it indicates that the appliance is effective in improving sleep breathing disorders.

[0136] The evaluation of snoring characteristics changes is to further verify the effect of the appliance by evaluating the changes in snoring characteristics; the snoring characteristics include snoring frequency, snoring duration, and snoring intensity; by comparing the changes in these indicators before and after the use of the appliance, combined with the evaluation of the AHI value, the system can provide a comprehensive evaluation result to judge the clinical effect of the appliance; ultimately, these evaluation data will be presented in the form of charts and reports to help doctors and patients intuitively understand the treatment effect and provide a basis for further treatment plans. The visual report generation module presents the effect evaluation results in the form of charts on the display of the smartphone.

[0137] The visualization report generation module is an AHI trend chart or a snoring spectrum chart or a calendar view or a statistical chart; the AHI trend chart shows the change trend of the patient's AHI value before and after the use of the corrective appliance; the snoring spectrum chart: shows the frequency distribution of snoring and its change trend; the calendar view: shows the frequency and intensity of snoring every day; the statistical charts include bar charts and line charts, which are used to show the changes in snoring evaluation indicators in different time periods.

[0138] The above-mentioned oral appliance is not only applicable to the APOA-A model oral appliance produced by the applicant, but also to all oral appliances on the market. The APOA-A model oral appliance product consists of a retainer and a grip (including a temperature label), and the retainer includes an upper tooth retainer and a lower tooth retainer. The grip is made of PP material, the upper and lower tooth retainers are made of EVA and PC, and the temperature sensing label is made of PET film.

[0139] See the following clinical examples:

[0140] We selected 2 patients with different conditions from many trial users, and each patient included data from one week before and one week after using the oral appliance. The following is an example of the process of using the oral appliance, combined with the results of the snoring detection system and the effect evaluation module:

[0141] Usage process: At night, the phone microphone is used to collect the sound of sleep. The distance between the phone and the human body is fixed to keep the data collected each time valid. The data is uploaded to the cloud. After noise filtering and signal enhancement, the system removes the noise in the room. MFCCs and deep learning models are used to identify the user's snoring events. The data includes data before and after using the appliance for comparison.

[0142] Detector 1:

[0143] Data before use: snoring frequency is 20 times per hour, AHI is 25, snoring intensity is 65 decibels, and the average snoring duration is 45 seconds.

[0144] Data after use: After one week of use, the snoring frequency was reduced to 10 times per hour, and AHI

[0145] Dropped to 15, the snoring intensity dropped to 55 decibels, and the average snoring duration was 30 seconds.

[0146] Visual report example (such as Figure 2 shown):

[0147] Sleep data overview of subject 1 (such as Figure 3 (shown)

[0148] The following table is a comparison of the tester 1 before and after the test

[0149] index Before use After use Number of snoring per hour 20 10 AHI 25 15 Snoring intensity (dB) 65 55 Snoring duration (sec) 45 30 Sleep quality score (out of 10 points) 5 8 Number of wake-ups 4 2 Sleep quality (%) 80 90

[0150] Detector 2:

[0151] Data before use: AHI was 31. The average snoring frequency was 25 times per night. The average snoring intensity was 74dB.

[0152] · Post-use data: The oral appliance was used continuously for two weeks. The snoring detection system collected his sleep data every day. The system data showed that Mr. Chen's snoring data fluctuated in the second week after using the appliance: the average AHI index in the second week was 28, with a small decrease. The average snoring frequency was 22 times, which was not much different from before use. The snoring intensity remained at around 70dB.

[0153] Effect evaluation: Although the snoring detection system recorded that the AHI index and snoring frequency of the tester 2 decreased slightly in the second week after using the appliance, the overall data did not change much. The snoring intensity had almost no significant improvement and remained in the high intensity range. After analysis by the effect evaluation module, it was concluded that the oral appliance was not effective in stopping Mr. Chen's snoring, and he might need to replace it with another type of appliance or combine it with other treatment options.

[0154] Visual report example (such as Figure 4 shown):

[0155] Sleep data overview of subject 2 (such as Figure 5 (shown)

[0156] The following table is a comparison of the tester 2 before and after the test:

[0157] index Before use After use Number of snoring per hour 31 29 AHI 25 23 Snoring intensity (dB) 74 72 Snoring duration (sec) 40 38 Sleep quality score (out of 10 points) 6 65 Number of wake-ups 3 2 Sleep quality (%) 80 82

[0158] The above are only some embodiments of the present invention. For those skilled in the art, several modifications and improvements can be made without departing from the inventive concept of the present invention, which all belong to the protection scope of the invention.

Claims

1. A snoring detection system applied to an oral appliance, characterized in that: The device comprises an oral appliance installed in the oral cavity of a tester, a smartphone installed outside the tester and wirelessly connected to the oral appliance, a snoring detection software program installed in the smartphone, and a cloud server wirelessly connected to the smartphone; The snoring detection software program includes an audio acquisition module connected to a microphone inside a smart phone and used to adopt audio data and connected to a cloud server to upload the audio data, a data preprocessing module provided in the smart phone and wirelessly connected to the cloud server and performing denoising and signal enhancement processing on the audio data after downloading, a feature extraction module provided in the smart phone and connected to the data preprocessing module and extracting the preprocessed audio signal and using it for snoring recognition and evaluation of key features, a snoring detection module provided in the smart phone and connected to the feature extraction module and accurately identifying snoring events in the audio data, an effect evaluation module provided in the smart phone and connected to the snoring detection module and to the oral appliance respectively and performing quantitative analysis and evaluation on the use effect of the tester, and a visual report generation module provided in the smart phone and connected to the effect evaluation module; The visualization report generation module presents the effect evaluation results in the form of charts on the display screen of the smart phone.

2. A snoring detection system for an oral appliance according to claim 1, characterized in that: The data preprocessing module includes a bandpass filter for removing high noise, and an adaptive filter connected to the bandpass filter; The bandpass filter removes noises higher than 8000 Hz and lower than 20 Hz, because snoring sounds are mainly concentrated between 20 Hz and 2000 Hz; the adaptive filter further removes environmental noise from the filtered signal.

3. A snoring detection system for oral appliances according to claim 2, characterized in that: The implementation of the adaptive filter is based on the LMS algorithm, which is used to estimate the parameters of the system to minimize the mean square error between the output signal and the expected signal. The formula is: w(n+1)=w(n)+2μe(n)x(n) Where: w(n) is the weight vector of the filter, μ is the step size factor used to control the rate of weight update, e(n) is the current error signal, and x(n) is the input signal; By iteratively updating the weights, the adaptive filter can adaptively remove noise and enhance the snoring signal.

4. A snoring detection system for oral appliances according to claim 3, characterized in that: The feature extraction module uses Mel frequency cepstral coefficients MFCCs; The calculation steps of the Mel frequency cepstral coefficients MFCCs are as follows: (1) The extraction process of MFCCs includes the following key steps: A. Pre-emphasis: First, the input audio signal is pre-emphasized to enhance the high-frequency components; then, the high-frequency attenuation caused by the damping effect of the vocal organs in the speech signal is offset; the pre-emphasis filter is a first-order high-pass filter, and its mathematical expression is: y(t)=x(t)-αx(t-1) Where: y(t) is the pre-emphasized signal, x(t) is the original input signal, and α is the pre-emphasis coefficient; B. Framing and windowing: The audio signal is non-stationary. Framing divides it into multiple short-term stable frames with 50% overlap between frames. On the basis of framing, a Hamming window is applied to each frame signal to reduce spectrum leakage. The expression is: Where: N is the length of the window, n is the current time point; C. Fast Fourier Transform: Perform fast Fourier transform on each frame of the windowed signal to convert it from the time domain to the frequency domain; the result of FFT is the power spectrum of the signal, which is expressed as: Where X(k) is the kth frequency component, x(n) is the input signal, and N is the number of FFT points; D. Mel frequency cepstrum coefficients: Use the Mel filter bank to weight the FFT result to generate the Mel spectrum; the Mel scale is a nonlinear frequency scale that conforms to the human auditory perception and is defined as: Where: m is the Mel frequency, f is the linear frequency (Hz); After converting the linear frequency to the Mel frequency, a set of triangular filters are applied on this scale to perform weighted summation on the power spectrum of the FFT to obtain the Mel spectrum coefficients; then, through logarithmic operations and discrete cosine transform (DCT), the Mel frequency cepstral coefficients (MFCCs) are obtained; DCT converts the features of the Mel spectrum from the frequency domain to the cepstral domain and compresses it into a set of low-dimensional feature vectors; MFCC is calculated as follows: Where: m represents the index of the cepstral coefficient, K is the number of Mel filter banks; (2) Feature optimization and dimensionality reduction In order to improve the distinguishing ability and computational efficiency of features, the principal component analysis (PCA) dimensionality reduction technique is also used to optimize the extracted MFCCs features; PCA projects high-dimensional feature vectors into low-dimensional space through linear transformation to retain the main information of the data; the optimization process of PCA is as follows: A. Calculate the covariance matrix: Evaluate the correlation between features by calculating the covariance matrix of MFCCs features; Where: C is the covariance matrix, N is the number of samples, x i is the feature vector of the i-th sample, μ is the mean of the feature vector; B. Eigenvalue decomposition: Perform eigenvalue decomposition on the covariance matrix to obtain eigenvalues ​​and corresponding eigenvectors; select the eigenvector corresponding to the maximum eigenvalue as the new feature space after dimensionality reduction; Cv i =λ i v i Where: i is the i-th eigenvalue, v i is the corresponding eigenvector; Finally, the MFCCs feature vector is projected into a new low-dimensional feature space through linear transformation to obtain the optimized feature vector.

5. The snoring detection system for oral appliances according to claim 1, characterized in that: The snoring detection module adopts a model based on a convolutional neural network, and the network structure can effectively capture the timing and frequency characteristics in the audio signal.

6. A snoring detection system for oral appliances according to claim 5, characterized in that: The model is based on the convolutional neural network. The input is the extracted MFCCs features and corresponding labels. After passing through multiple convolutional layers and pooling layers, a high-dimensional feature map is extracted. The process is as follows: (1) The extracted MFCCs features are firstly subjected to spatial feature extraction through multiple convolutional layers; assuming that the output of the lth convolutional layer is H (l) , then it is calculated by the following formula: H (l) =σ(W (l) *H (l-1) +b (l) ) Among them, H (l) is the lth layer of convolution, * indicates the convolution operation, b (l) is the bias term, σ is the activation function (ReLU); (2) After each convolutional layer, a pooling layer is applied to reduce the dimension of the feature map and extract the main features; the commonly used pooling operation is maximum pooling, and its calculation formula is: Among them, R is the pooling window area; After multiple layers of convolution and pooling, a high-dimensional feature map is obtained. The feature map is flattened and input into the fully connected layer, and finally a sigmoid activation function is used to output the probability value of each time segment containing a snoring event. Among them, h is the flattened eigenvector, W FC and b FC are the weights and biases of the fully connected layer respectively; The network loss function uses the binary cross entropy loss function: Where: N is the total number of samples, y i is the true label (0 or 1) of the i-th sample, p i is the probability of snoring predicted by the model; The model is trained through back-propagation and stochastic gradient descent algorithms to minimize the loss function and optimize the model parameters.

7. The snoring detection system for oral appliances according to claim 5, characterized in that: The snoring detection module is also provided with a model optimization function; The model optimization function is that the user evaluates the detection results through the feedback interface provided by the system, generates a feedback data set, combines the feedback data with the original training data to form an extended training set, uses the extended training set to retrain the model, and further adjusts the model parameters through the optimization algorithm Adam optimizer to adapt to the new data distribution and user needs; regularly evaluates the performance of the optimized model to ensure the detection effect of the model in different environments and conditions.

8. The snoring detection system for oral appliances according to claim 1, characterized in that: The effect evaluation module includes AHI evaluation and snoring characteristic change evaluation; The AHI assessment is an important indicator for measuring the severity of sleep apnea, which is defined as the sum of the number of apneas and hypopneas occurring per hour of sleep; The specific calculation is disclosed as follows: The effect evaluation module detects apnea and hypopnea events in the signal processing and feature extraction module; the start and end time of each event is marked by analyzing the energy change, breathing pattern and audio feature MFCCs of the audio signal; the AHI value is calculated based on the number of apnea and hypopnea events detected and the total sleep time of the patient; in order to improve the accuracy of the calculation, the AHI is calculated within a specific time window (every hour), and then the AHI in the entire sleep cycle is averaged. By comparing the AHI values ​​before and after the use of the appliance, the therapeutic effect of the appliance is evaluated; if the AHI value is significantly reduced after the use of the appliance, especially when the AHI drops to a normal range during sleep (AHI<5), it indicates that the appliance is effective in improving sleep breathing disorders; The snoring characteristic change evaluation further verifies the effect of the appliance by evaluating the changes in snoring characteristics; the snoring characteristics include snoring frequency, snoring duration, and snoring intensity; by comparing the changes in these indicators before and after the use of the appliance, combined with the evaluation of the AHI value, the system can provide a comprehensive evaluation result for judging the clinical effect of the appliance; ultimately, these evaluation data will be presented in the form of charts and reports to help doctors and patients intuitively understand the treatment effect and provide a basis for further treatment plans.

9. The snoring detection system for oral appliances according to claim 1, characterized in that: The visual report generation module is an AHI trend graph or a snoring spectrum graph or a calendar view or a statistical chart; The AHI trend graph shows the change trend of the patient's AHI value before and after the use of the appliance; The snoring spectrum diagram: shows the frequency distribution of snoring and its changing trend; The calendar view: displays the frequency and intensity of snoring every day; The statistical charts include bar charts and line charts, which are used to show the changes in snoring evaluation indicators in different time periods.

Images