Obstructive sleep apnea judgment method based on artificial intelligence

Abstract

The invention discloses an artificial intelligence-based obstructive sleep apnea judgment method, which comprises the following steps of: S1, acquiring the respiratory rate, the heart rate and the oxygen saturation of a user in a sleep period, and constructing a physiological data set; s2, filtering, denoising and cleaning the collected original physiological data set to remove noise and abnormal values; s3, key physiological features are extracted from the preprocessed physiological data set; s4, training a judgment model by utilizing improved deep learning, and performing comprehensive analysis by the judgment model according to the extracted physiological features to judge whether obstructive sleep apnea occurs or not; s5, the system monitors the sleep state of the user in real time, and provides instant feedback or alarm when obstructive sleep apnea is detected; and S6, the system performs analysis according to the physiological features of different users, and provides judgment results and health suggestions for individuals. The accuracy and real-time performance of obstructive sleep apnea judgment are improved, and accurate sleep apnea monitoring and judgment are achieved.

Description

Technical Field

[0001] The present invention relates to the field of sleep breathing technology, and in particular to an obstructive sleep apnea determination method based on artificial intelligence. Background Art

[0002] Obstructive sleep apnea syndrome is one of the most common sleep disorders, which is mainly manifested by repeated apnea or hypopnea during sleep. This syndrome not only seriously affects the patient's sleep quality, but also leads to the occurrence of multi-system diseases such as cancer, hypertension, diabetes, stroke, cognitive impairment, etc., and is considered to be one of the serious public health problems. Therefore, timely and accurate detection and determination of obstructive sleep apnea in the early stage is of great significance to improve the health of patients.

[0003] At present, the traditional method of determining obstructive sleep apnea mainly relies on polysomnography, which determines whether sleep apnea exists by monitoring and analyzing multiple physiological parameters of the patient during sleep (such as respiratory rate, heart rate, oxygen saturation, electroencephalogram, etc.). However, the polysomnography process is complicated and needs to be performed in a professional sleep center. Patients need to wear multiple sensors, which not only interferes with the patient's sleep, but also increases the cost and inconvenience of detection. It has not been popularized in my country. In addition, traditional methods often rely on a single physiological parameter and a simple algorithm, resulting in insufficient accuracy and reliability of the determination results, making it difficult to timely reflect the user's actual sleep status.

[0004] In traditional methods, commonly used algorithms such as simple threshold judgment methods and statistical analysis methods cannot fully utilize the information of multiple physiological parameters and have poor adaptability to individual differences. For example, physiological characteristics such as respiratory rate, heart rate, and oxygen saturation fluctuations vary significantly among different individuals. A single threshold judgment method is difficult to adapt to the needs of different users, resulting in serious misjudgments and missed judgments. At the same time, traditional methods usually lack real-time performance and cannot provide sleep apnea judgment results in a timely manner, making it difficult to meet users' needs for instant feedback and health management.

[0005] In recent years, artificial intelligence and machine learning technologies have been widely used in the medical field, providing new ideas and methods for the detection and determination of obstructive sleep apnea. By combining multiple physiological parameters and using advanced technologies such as deep learning, graph neural networks, and multimodal fusion, efficient, accurate, and precise determination of sleep apnea can be achieved. However, there are still some shortcomings in current research and applications. For example, the existing artificial intelligence models have not fully combined the multimodal characteristics of different physiological parameters in feature extraction and model construction, resulting in the accuracy and real-time nature of the determination results still need to be improved; in addition, there is little research on accurate determination, and it is difficult to provide targeted health advice and improvement measures based on the physiological characteristics of different users. Summary of the invention

[0006] The embodiment of the present invention proposes an obstructive sleep apnea determination method based on artificial intelligence, which can improve the accuracy and real-time performance of obstructive sleep apnea determination and achieve accurate sleep apnea monitoring and determination.

[0007] According to an embodiment of the present invention, a method for determining obstructive sleep apnea based on artificial intelligence is proposed, comprising the following steps:

[0008] S1. Use non-invasive biosensors to collect the user's breathing rate, heart rate and oxygen saturation in real time during sleep to build a physiological data set;

[0009] S2, filtering, denoising and cleaning the collected original physiological data set to remove noise and outliers;

[0010] S3, extracting key physiological features from the preprocessed physiological data set, wherein the physiological features include respiratory rate, heart rate variability, and oxygen saturation fluctuation;

[0011] S4. Using improved deep learning to train a determination model, wherein the determination model performs a comprehensive analysis based on the extracted physiological features to determine whether sleep apnea occurs;

[0012] S5. The system monitors the user's sleep status in real time and provides immediate feedback or alarm when sleep apnea is detected;

[0013] S6. The system analyzes the physiological characteristics of different users and provides individual judgment results and health recommendations.

[0014] Optionally, the S1 specifically includes:

[0015] S11, arranging a plurality of non-invasive biosensors in the user's sleeping environment to collect the user's physiological parameter data during sleep in real time, wherein the physiological parameter data includes the respiratory rate R f , heart rate r and oxygen saturation O s ;

[0016] S12, the non-invasive biosensor transmits the collected respiratory rate R through the wireless communication module f , heart rate r and oxygen saturation O s The data is transmitted to the data processing terminal, which timestamps each data point and constructs a time series physiological data set D raw :

[0017] D ra w={(t1,R f(t1),H r (t1),O s (t1)),(t2,R f (t2),H r (t2),O s (t2)),…,(t n ,R f (t n ),H r (t n ),O s (t n ))}

[0018] Among them, t i represents the i-th time point, R f (t i ), H r (t i ), O s (t i ) represent the time point t i Respiratory rate, heart rate and oxygen saturation were collected.

[0019] Optionally, the S2 specifically includes:

[0020] S21. Constructing the time series physiological data set D raw The respiratory rate R f , heart rate r and oxygen saturation O s The data was filtered and a bandpass filter was used to remove high-frequency and low-frequency noise. The filtered time series physiological data set D filtered ;

[0021] S22, for the filtered time series physiological data set D filtered The data set was denoised and multi-scale analysis was performed on the data using wavelet transform to remove background noise and instantaneous interference. The denoised time series physiological data set D denoised ;

[0022] S23. Denoised time series physiological data set D denoised The dataset was cleaned to remove outliers and incomplete data, and statistical methods were used to identify and replace abnormal data points. The cleaned time series physiological dataset D clean .

[0023] Optionally, the S3 specifically includes:

[0024] S31. From the time series physiological dataset D clean Extract comprehensive physiological features from the respiratory rate R f , heart rate variability HRV and oxygen saturation Os Perform multi-parameter fusion to generate a preliminary comprehensive feature vector, calculate the comprehensive physiological feature vector at each time point in the preliminary comprehensive feature vector, including the time domain and frequency domain characteristics of respiratory frequency, multi-scale entropy of heart rate variability, and high-order statistics of oxygen saturation fluctuations, and generate the final comprehensive feature vector F combined ;

[0025] S32, combined with multi-parameter fusion technology, use principal component analysis to analyze the comprehensive physiological feature vector F combined Perform dimensionality reduction to obtain the main characteristic component F PCA , the main feature component F after dimensionality reduction PCA As input features, the final feature set for sleep apnea determination is constructed.

[0026] Optionally, the S31 specifically includes:

[0027] S311. From the time series physiological data set D clean Extract the respiratory rate R f , heart rate variability HRV and oxygen saturation O s Nonlinear features of the proposed algorithm are used to construct a preliminary feature set.

[0028] S312, respiratory rate R f The radial basis function kernel method is used to capture nonlinear features and obtain the respiratory frequency feature F Rf ;

[0029]

[0030] Among them, α j is the weight of the kernel method, σ is the parameter of the basis function, and N is the number of samples;

[0031] S313, extracting the heart rate variability HRV by kernel method, using polynomial kernel function to capture nonlinear features, and obtaining the heart rate variability feature F HRV :

[0032]

[0033] Where c is a constant and d is the degree of the polynomial kernel;

[0034] S314. Perform kernel method feature extraction on oxygen saturation fluctuation Os, and use Sigmoid kernel function to capture nonlinear features to obtain oxygen saturation feature F Os :

[0035] F Os =tanh(β(O s (t i )·O s (tj ))+c);

[0036] Among them, β and c are the parameters of the Sigmoid kernel;

[0037] S315, using improved variational autoencoder to process and fuse respiratory frequency features F Rf , heart rate variability characteristics F HRV and oxygen saturation characteristics F Os , generate the comprehensive feature vector F combined , the loss function of VAE combines sleep breathing data, including reconstruction loss and regularization terms of physiological data features, and is defined as:

[0038]

[0039] Among them, x represents the input physiological data feature, z is the latent variable, φ and θ are the parameters of VAE, KL represents the Kullback-Leibler divergence, T i is the time series length of the i-th sample, They represent the reconstructed respiratory rate, heart rate variability and oxygen saturation respectively, λ1 and λ2 are the weights of the regularization term;

[0040] S316, the final comprehensive feature vector F generated combined :

[0041] F combined =[F Rf ,F HRV ,F Os ].

[0042] Optionally, the S32 specifically includes:

[0043] S321, the comprehensive physiological feature vector F combined All data are standardized, the mean and standard deviation of each feature are calculated, and the feature vector is normalized to zero mean and unit variance. standard ;

[0044] S322, the standardized feature vector F standard Perform principal component analysis, calculate the characteristic covariance matrix Σ, and solve the eigenvalues ​​and eigenvectors;

[0045]

[0046] Among them, μ represents the mean of the standardized feature, and n is the number of samples;

[0047] S323, select the first k main characteristic components, sort them according to the size of the eigenvalues, select the first k eigenvectors, and construct the main characteristic component matrix W;

[0048] W=[w1,w2,…,w k ];

[0049] Among them, w i represents the i-th main eigenvector, k is the number of main characteristic components selected;

[0050] S324, the standardized feature vector F standard Projected onto the main characteristic component matrix W, the main characteristic component F after dimensionality reduction is obtained PCA ;

[0051] F PCA (t i )=F standard (t i )·W;

[0052] S325, the main feature component F after dimensionality reduction PCA As input features, the final feature set for sleep apnea determination is constructed.

[0053] Optionally, the S4 specifically includes:

[0054] S41, using the preprocessed comprehensive feature vector F combined As input features, a deep learning model combining multimodal fusion and graph neural network is constructed for comprehensive analysis and determination of sleep apnea;

[0055] S42, through multimodal fusion technology, the comprehensive feature vector F combined Processing, integrating the spatial and temporal information of different physiological signals, generating a multimodal feature representation F multi ;

[0056] F multi =Fusion(F Rf ,F HRV ,F Os );

[0057] Among them, Fusion represents the multimodal fusion function, combined with the respiratory rate R f , heart rate variability HRV and oxygen saturation O s feature;

[0058] S43, multimodal feature representation F multi It is constructed as a heterogeneous graph, where nodes represent physiological characteristics at different time points, edges represent the correlation of time series, and different weights are assigned according to sleep breathing patterns;

[0059] S44. Use graph neural networks to process heterogeneous graphs, extract complex graph structure features, and generate high-dimensional feature representations F GNN ;

[0060] F GNN =GNN(F multi );

[0061] S45. Pre-train GNN through self-supervised learning technology to construct the sleep apnea judgment task to maximize the model's ability to capture potential patterns;

[0062]

[0063] Among them, y i is the actual label, the presence or absence of sleep apnea, p(y i ) is the predicted probability;

[0064] S46. Combine contrastive learning strategies to pair samples with similar features and optimize the judgment accuracy of the model;

[0065]

[0066] Among them, sim represents the similarity function, which is used to calculate the high-dimensional feature representation F GNN The similarity between them, τ is the temperature parameter, which is used to adjust the sensitivity of contrastive learning;

[0067] S47, after the training is completed, the optimized deep learning model is used to determine sleep apnea in real time, and the input comprehensive feature vector F combined Pass it to the model to generate the final judgment result;

[0068] S48. During the model determination process, a final determination result is generated:

[0069] R final =Decoder(F GNN );

[0070] Among them, Decoder is the decoder part of the model, R final This is the final determination result, indicating whether the user has sleep apnea.

[0071] Optionally, the S48 specifically includes:

[0072] S481. The high-dimensional feature representation F generated from the graph neural network GNN Input into the decoder model, which includes a multi-layer fully connected neural network;

[0073] S482, in the first layer of the decoder, the high-dimensional feature representation F of the inputGNN Perform linear transformation to generate intermediate feature representation F hidden ;

[0074] S483. In the second layer of the decoder, the intermediate feature representation F hidden Perform nonlinear transformation to generate output feature representation F output ;

[0075] S484, in the output layer of the decoder, the output feature representation F output Perform classification operations to generate the final judgment result R final ;

[0076] S485, in the process of determining the model, combining the threshold setting to determine whether the user has sleep apnea, when R final When the probability value in is greater than the preset threshold, it is determined that sleep apnea occurs;

[0077]

[0078] S486, the final judgment result R final And the corresponding probability values ​​are stored in the database

[0079] Optionally, the S6 specifically includes:

[0080] S61: construct a user feature database based on the user's historical physiological data and current physiological data, wherein the user feature includes a respiratory rate R f , heart rate variability HRV and oxygen saturation O s ;

[0081]

[0082] in, HRV i , They represent the respiratory rate, heart rate variability and oxygen saturation characteristics of the i-th user, respectively, and N is the number of users;

[0083] S62, using a clustering algorithm to perform cluster analysis on the data in the user feature database, identifying feature patterns of different user groups, and generating a user feature clustering model;

[0084]

[0085] Among them, γ ik is an indicator variable. When the i-th user belongs to the k-th cluster, γ ik =1, otherwise γ ik =0;μ k is the center of the kth cluster;

[0086] S63, matching the physiological characteristics of the current user with the user characteristic clustering model to determine the characteristic group to which the current user belongs;

[0087]

[0088] in, HRV current and Respectively represent the current user's breathing rate, heart rate variability and oxygen saturation characteristics;

[0089] S64, generating a determination result and health advice based on the matching result and the specific physiological characteristics of the current user, wherein the determination result includes whether sleep apnea occurs, and the health advice includes specific measures to improve sleep quality;

[0090] S65, using a recommendation algorithm to provide health advice to the user based on the user's physiological characteristics and the determination result;

[0091] S66, feeding back the determination results and health suggestions to the user in real time through the user terminal device, providing guidance for the user's health management and sleep quality improvement;

[0092] S67. With the user's consent, the user's determination results and health recommendations are stored in a user feature database, and the determination model and recommendation algorithm are optimized.

[0093] Beneficial effects of the present invention:

[0094] (1) In the feature extraction stage, the present invention uses non-invasive biosensors to collect multiple physiological parameters of the user, such as respiratory rate, heart rate variability and oxygen saturation in real time, and uses improved radial basis function kernel method, polynomial kernel function and sigmoid kernel function to extract nonlinear features, capture the complex nonlinear relationship of physiological signals, and generate more accurate and comprehensive feature representation. After fusion and dimensionality reduction processing of variational autoencoders, the feature representation ability and the performance of the judgment model are further improved.

[0095] (2) During the model building stage, the present invention uses multimodal fusion technology to fuse the spatial and temporal information of respiratory frequency, heart rate variability and oxygen saturation features to generate a multimodal feature representation. It uses graph neural networks to construct heterogeneous graphs, extract complex graph structure features, and generate high-dimensional feature representations. The graph neural network can fully capture the correlation and potential patterns between physiological signals, thereby improving the accuracy of sleep apnea judgment.

[0096] (3) The present invention uses clustering algorithms and recommendation algorithms to identify the characteristic patterns of different user groups based on the user's historical physiological data and current physiological data, generate a user characteristic clustering model, and determine the characteristic group to which the user belongs by matching the current user's physiological characteristics with the clustering model, and generate judgment results and health recommendations. The analysis is combined with the user's specific physiological characteristics to provide targeted sleep quality improvement measures, thereby improving the applicability of the system and user experience.

[0097] (4) In the process of model training and judgment, the present invention combines self-supervised learning and contrastive learning strategies to maximize the model's ability to capture potential patterns, and optimizes the model's judgment accuracy through self-supervised pre-training and contrastive learning. After training, the optimized deep learning model can monitor the user's sleep state in real time and provide immediate feedback or alarms when sleep apnea is detected. The real-time function ensures that users can be reminded in time during sleep to prevent potential health risks. BRIEF DESCRIPTION OF THE DRAWINGS

[0098] The accompanying drawings are used to provide a further understanding of the present invention and constitute a part of the specification. Together with the embodiments of the present invention, they are used to explain the present invention and do not constitute a limitation of the present invention. In the accompanying drawings:

[0099] Figure 1 The present invention provides a flowchart of an obstructive sleep apnea determination method based on artificial intelligence. DETAILED DESCRIPTION

[0100] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, which only illustrate the basic structure of the present invention in a schematic manner, and therefore only show the components related to the present invention.

[0101] refer to Figure 1 , an obstructive sleep apnea determination method based on artificial intelligence, comprising the following steps:

[0102] S1. Use non-invasive biosensors to collect the user's breathing rate, heart rate and oxygen saturation in real time during sleep to build a physiological data set;

[0103] S2, filtering, denoising and cleaning the collected original physiological data set to remove noise and outliers;

[0104] S3, extracting key physiological features from the preprocessed physiological data set, the physiological features including respiratory rate, heart rate variability, and oxygen saturation fluctuation;

[0105] S4. Using improved deep learning to train a determination model, the determination model performs a comprehensive analysis based on the extracted physiological features to determine whether sleep apnea occurs;

[0106] S5. The system monitors the user's sleep status in real time and provides immediate feedback or alarm when sleep apnea is detected;

[0107] S6. The system analyzes the physiological characteristics of different users and provides individual judgment results and health recommendations.

[0108] In this implementation, S1 specifically includes:

[0109] S11. Arrange multiple non-invasive biosensors in the user's sleeping environment to collect the user's physiological parameter data during sleep in real time. The physiological parameter data includes the respiratory rate R f , heart rate r and oxygen saturation O s ;

[0110] S12, the non-invasive biosensor transmits the collected respiratory rate R through the wireless communication module f , heart rate r and oxygen saturation O s The data is transmitted to the data processing terminal, which timestamps each data point and constructs a time series physiological data set D. raw :

[0111] D raw ={(t1,R f (t1),H r (t1),O s (t1)),(t2,R f (t2),H r (t2),O s (t2)),…,(t n ,R f (t n ),H r (t n ),O s (t n ))};

[0112] Among them, t i represents the i-th time point, R f (t i ), H r (t i ), O s (t i ) represent the time point t i Respiratory rate, heart rate and oxygen saturation were collected.

[0113] In this implementation, S2 specifically includes:

[0114] S21. Constructing the time series physiological data set Draw The respiratory rate R f , heart rate r and oxygen saturation O s The data was filtered and a bandpass filter was used to remove high-frequency and low-frequency noise. The filtered time series physiological data set D filtered ;

[0115] S22, for the filtered time series physiological data set D filtered The data set was denoised and multi-scale analysis was performed on the data using wavelet transform to remove background noise and instantaneous interference. The denoised time series physiological data set D denoised ;

[0116] S23. Denoised time series physiological data set D denoised The dataset was cleaned to remove outliers and incomplete data, and statistical methods were used to identify and replace abnormal data points. The cleaned time series physiological dataset D clean .

[0117] In this implementation, S3 specifically includes:

[0118] S31. From the time series physiological dataset D clean Extract comprehensive physiological features from the respiratory rate R f , heart rate variability HRV and oxygen saturation O s Perform multi-parameter fusion to generate a preliminary comprehensive feature vector, calculate the comprehensive physiological feature vector at each time point in the preliminary comprehensive feature vector, including the time domain and frequency domain characteristics of respiratory frequency, multi-scale entropy of heart rate variability, and high-order statistics of oxygen saturation fluctuations, and generate the final comprehensive feature vector F combined ;

[0119] S32, combined with multi-parameter fusion technology, use principal component analysis to analyze the comprehensive physiological feature vector F combined Perform dimensionality reduction to obtain the main characteristic component F PCA , the main feature component F after dimensionality reduction PCA As input features, the final feature set for sleep apnea determination is constructed.

[0120] In this implementation, S31 specifically includes:

[0121] S311. From the time series physiological data set D clean Extract the respiratory rate R f , heart rate variability HRV and oxygen saturation O s Nonlinear features of the proposed algorithm are used to construct a preliminary feature set.

[0122] S312, respiratory rate R fThe radial basis function kernel method is used to capture nonlinear features and obtain the respiratory frequency feature F Rf ;

[0123]

[0124] Among them, α j is the weight of the kernel method, σ is the parameter of the basis function, and N is the number of samples;

[0125] S313, extracting the heart rate variability HRV by kernel method, using polynomial kernel function to capture nonlinear features, and obtaining the heart rate variability feature F HRV :

[0126]

[0127] Where c is a constant and d is the degree of the polynomial kernel;

[0128] S314. Perform kernel method feature extraction on oxygen saturation fluctuation Os, and use Sigmoid kernel function to capture nonlinear features to obtain oxygen saturation feature F Os :

[0129]

[0130] Among them, β and c are the parameters of the Sigmoid kernel;

[0131] S315, using improved variational autoencoder to process and fuse respiratory frequency features F Rf , heart rate variability characteristics F HRV and oxygen saturation characteristics F Os , generate the comprehensive feature vector F combined , the loss function of VAE combines sleep breathing data, including reconstruction loss and regularization terms of physiological data features, and is defined as:

[0132]

[0133] Among them, x represents the input physiological data feature, z is the latent variable, φ and θ are the parameters of VAE, KL represents the Kullback-Leibler divergence, T i is the time series length of the i-th sample, They represent the reconstructed respiratory rate, heart rate variability and oxygen saturation respectively, λ1 and λ2 are the weights of the regularization term;

[0134] S316, the final comprehensive feature vector F generated combined :

[0135] F combined =[F Rf ,FHRV ,F Os ].

[0136] In this implementation, S32 specifically includes:

[0137] S321, the comprehensive physiological feature vector F combined All data are standardized, the mean and standard deviation of each feature are calculated, and the feature vector is normalized to zero mean and unit variance. standard ;

[0138] S322, the standardized feature vector F standard Perform principal component analysis, calculate the characteristic covariance matrix Σ, and solve the eigenvalues ​​and eigenvectors;

[0139]

[0140] Among them, μ represents the mean of the standardized feature, and n is the number of samples;

[0141] S323, select the first k main characteristic components, sort them according to the size of the eigenvalues, select the first k eigenvectors, and construct the main characteristic component matrix W;

[0142] W=[w1,w2,…,w k ];

[0143] Among them, w i represents the i-th main eigenvector, k is the number of main characteristic components selected;

[0144] S324, the standardized feature vector F standard Projected onto the main characteristic component matrix W, the main characteristic component F after dimensionality reduction is obtained PCA ;

[0145] F PCA (t i )=F standard (t i )·W;

[0146] S325, the main feature component F after dimensionality reduction PCA As input features, the final feature set for sleep apnea determination is constructed.

[0147] In this implementation, S4 specifically includes:

[0148] S41, using the preprocessed comprehensive feature vector F combined As input features, a deep learning model combining multimodal fusion and graph neural network is constructed for comprehensive analysis and determination of sleep apnea;

[0149] S42, through multimodal fusion technology, the comprehensive feature vector F combined Processing, integrating the spatial and temporal information of different physiological signals, generating a multimodal feature representation F multi ;

[0150] F multi =Fusion(F Rf ,F HRV ,F Os );

[0151] Among them, Fusion represents the multimodal fusion function, combined with the respiratory rate R f , heart rate variability HRV and oxygen saturation O s feature;

[0152] S43, multimodal feature representation F multi It is constructed as a heterogeneous graph, where nodes represent physiological characteristics at different time points, edges represent the correlation of time series, and different weights are assigned according to sleep breathing patterns;

[0153] S44. Use graph neural networks to process heterogeneous graphs, extract complex graph structure features, and generate high-dimensional feature representations F GNN ;

[0154] F GNN =GNN(F multi );

[0155] S45. Pre-train GNN through self-supervised learning technology to construct the sleep apnea judgment task to maximize the model's ability to capture potential patterns;

[0156]

[0157] Among them, y i is the actual label, the presence or absence of sleep apnea, p(y i ) is the predicted probability;

[0158] S46. Combine contrastive learning strategies to pair samples with similar features and optimize the judgment accuracy of the model;

[0159]

[0160] Among them, sim represents the similarity function, which is used to calculate the high-dimensional feature representation F GNN The similarity between them, τ is the temperature parameter, which is used to adjust the sensitivity of contrastive learning;

[0161] S47, after the training is completed, the optimized deep learning model is used to determine sleep apnea in real time, and the input comprehensive feature vector F combinedPass it to the model to generate the final judgment result;

[0162] S48. During the model determination process, a final determination result is generated:

[0163] R final =Decoder(F GNN );

[0164] Among them, Decoder is the decoder part of the model, R final This is the final determination result, indicating whether the user has sleep apnea.

[0165] In this implementation, S48 specifically includes:

[0166] S481. The high-dimensional feature representation F generated from the graph neural network GNN Input into the decoder model, which includes a multi-layer fully connected neural network;

[0167] S482, in the first layer of the decoder, the high-dimensional feature representation F of the input GNN Perform linear transformation to generate intermediate feature representation F hidden ;

[0168] S483. In the second layer of the decoder, the intermediate feature representation F hidden Perform nonlinear transformation to generate output feature representation F output ;

[0169] S484, in the output layer of the decoder, the output feature representation F output Perform classification operations to generate the final judgment result R final ;

[0170] S485, in the process of determining the model, combining the threshold setting to determine whether the user has sleep apnea, when R final When the probability value in is greater than the preset threshold, it is determined that sleep apnea occurs;

[0171]

[0172] S486, the final judgment result R final And the corresponding probability values ​​are stored in the database

[0173] In this implementation, S6 specifically includes:

[0174] S61: Build a user feature database based on the user's historical physiological data and current physiological data. The user features include respiratory rate R f , heart rate variability HRV and oxygen saturation O s ;

[0175]

[0176] in, HRV i , They represent the respiratory rate, heart rate variability and oxygen saturation characteristics of the i-th user, respectively, and N is the number of users;

[0177] S62, using a clustering algorithm to perform cluster analysis on the data in the user feature database, identifying feature patterns of different user groups, and generating a user feature clustering model;

[0178]

[0179] Among them, γ ik is an indicator variable. When the i-th user belongs to the k-th cluster, γ ik =1, otherwise γ ik =0;μ k is the center of the kth cluster;

[0180] S63, matching the physiological characteristics of the current user with the user characteristic clustering model to determine the characteristic group to which the current user belongs;

[0181]

[0182] in, HRV current and Respectively represent the current user's breathing rate, heart rate variability and oxygen saturation characteristics;

[0183] S64, generating a determination result and health advice based on the matching result and the specific physiological characteristics of the current user, wherein the determination result includes whether sleep apnea occurs, and the health advice includes specific measures to improve sleep quality;

[0184] S65, using a recommendation algorithm to provide health advice to the user based on the user's physiological characteristics and the determination result;

[0185] S66, feeding back the determination results and health suggestions to the user in real time through the user terminal device, providing guidance for the user's health management and sleep quality improvement;

[0186] S67. With the user's consent, the user's determination results and health recommendations are stored in a user feature database, and the determination model and recommendation algorithm are optimized.

[0187] Embodiment 1:

[0188] In a sleep center of a large hospital, in order to improve the diagnosis and treatment of sleep apnea patients, researchers decided to apply the method of the present invention to test 30 patients. These patients were between 30 and 60 years old and had varying degrees of sleep apnea symptoms. The traditional polysomnography detection method was used to compare with the method of the present invention to verify the effectiveness and advantages of the present invention.

[0189] Before falling asleep at 9 pm, the patient wears non-invasive biosensors, including respiratory sensors, heart rate sensors, and oxygen saturation sensors. The sensors transmit the collected data to the data processing terminal in real time through the wireless communication module. The data processing terminal filters, denoises, and cleans the collected respiratory rate, heart rate variability, and oxygen saturation data to remove noise and outliers, and extract key physiological features. The extracted features are processed through multimodal fusion technology and graph neural networks to generate high-dimensional feature representations, and ultimately determine whether the patient has sleep apnea.

[0190] In a specific sleep apnea determination result, taking patient A as an example:

[0191] During the sleep of patient A, the sensor recorded his breathing rate, heart rate variability and oxygen saturation data per minute. After data preprocessing and feature extraction, the system generated a comprehensive physiological feature vector of patient A. After multimodal fusion and graph neural network processing, the system determined that patient A had 23 sleep apnea events during the entire sleep process. The judgment results showed that patient A's breathing rate dropped sharply when the sleep apnea event occurred, the heart rate variability increased significantly, and the oxygen saturation fluctuated violently.

[0192] In this test, the researchers collected data from 30 patients and compared the results of the method of the present invention with the results of the traditional polysomnography detection method. The specific data are shown in Table 1 below:

[0193] Table 1 Comparison test data of sleep apnea determination methods

[0194]

[0195] In comparative tests, the method of the present invention is very close to the traditional method in determining the number of sleep apnea events, with an average error of only 5.2%. In addition, the method of the present invention also shows significant advantages in data processing speed and patient experience. Traditional polysomnography needs to be performed in a sleep center, and patients need to wear multiple sensors, which affects sleep comfort; the method of the present invention uses non-invasive biosensors, which does not affect the patient's normal sleep, and the data processing and determination process is carried out in real time during the patient's sleep, and can provide determination results and health advice in a timely manner.

[0196] Specific data include:

[0197] In a test on a certain night, the data comparison between the method of the present invention and the traditional method is as follows:

[0198] Time: June 15, 2024, Location: Sleep Center of a large hospital

[0199] Patient A: The number of apnea determined by the traditional method was 25 times, while the number determined by the method of the present invention was 23 times, with an error rate of 8%

[0200] Patient B: The number of apnea determined by the traditional method was 19 times, while the method of the present invention determined it to be 18 times, with an error rate of 5.3%

[0201] Patient C: The number of apnea determined by the traditional method was 29 times, while the number determined by the method of the present invention was 30 times, with an error rate of 3.4%

[0202] Patient D: The number of apnea determined by the traditional method was 16 times, while the method of the present invention determined it to be 15 times, with an error rate of 6.25%

[0203] Patient E: The number of apnea determined by the traditional method was 21 times, while the number determined by the method of the present invention was 22 times, with an error rate of 4.76%

[0204] The data show that the method of the present invention is comparable to traditional methods in terms of accuracy and real-time performance, and even outperforms them in some aspects. In addition, by combining multimodal feature fusion and graph neural networks, the method of the present invention can more comprehensively analyze and determine obstructive sleep apnea events, provide health advice to patients, and improve the applicability of the system and user experience.

[0205] Table 2 Guidance data for patient severity classification based on the number of apneas

[0206]

[0207] The method of the present invention classifies sleep apnea into mild, moderate and severe based on the number of sleep apnea events. The classification is based on the number of apneas per hour, which directly reflects the degree and frequency of the patient's respiratory restriction. Frequent apneas can lead to insufficient oxygen supply, interfere with sleep structure, and further affect overall health. Therefore, based on the classification, medical professionals can accurately and objectively judge the severity of the disease, thereby formulating a more appropriate treatment plan, such as CPAP treatment or surgical intervention.

[0208] In summary, this embodiment fully demonstrates the significant advantages of the method of the present invention in solving the problems of the prior art by describing in detail the application process and data comparison in specific scenarios, and reflects its practicality and innovation in sleep apnea determination.

[0209] In the feature extraction stage, the present invention uses non-invasive biosensors to collect the user's respiratory rate, heart rate variability, oxygen saturation and other physiological parameters in real time, and uses improved radial basis function kernel method, polynomial kernel function and Sigmoid kernel function to extract nonlinear features, capture the complex nonlinear relationship of physiological signals, and generate more accurate and comprehensive feature representation. After fusion and dimensionality reduction processing of variational autoencoders, the feature representation ability and the performance of the judgment model are further improved.

[0210] In the model building stage, the present invention uses multimodal fusion technology to fuse the spatial and temporal information of respiratory frequency, heart rate variability and oxygen saturation features to generate multimodal feature representation, and uses graph neural network to construct heterogeneous graphs, extract complex graph structure features, and generate high-dimensional feature representation. The graph neural network can fully capture the correlation and potential patterns between physiological signals, thereby improving the accuracy of determining obstructive sleep apnea.

[0211] The present invention uses clustering algorithms and recommendation algorithms to identify the characteristic patterns of different user groups based on the user's historical and current physiological data, generate a user characteristic clustering model, and determine the characteristic group to which the user belongs by matching the current user's physiological characteristics with the clustering model, and generate judgment results and health recommendations. The analysis combines the user's specific physiological characteristics to provide targeted sleep quality improvement measures, thereby improving the system's applicability and user experience.

[0212] In the process of model training and judgment, the present invention combines self-supervised learning and contrastive learning strategies to maximize the model's ability to capture potential patterns, and optimizes the model's judgment accuracy through self-supervised pre-training and contrastive learning. After training, the optimized deep learning model can monitor the user's sleep state in real time and provide immediate feedback or alarms when sleep apnea is detected. The real-time function ensures that users can be reminded in time during sleep to prevent potential health risks.

[0213] The above description is only a preferred specific implementation manner of the present invention, but the protection scope of the present invention is not limited thereto. Any technician familiar with the technical field can make equivalent replacements or changes according to the technical scheme and inventive concept of the present invention within the technical scope disclosed by the present invention, which should be covered by the protection scope of the present invention.

Claims

1. An obstructive sleep apnea determination method based on artificial intelligence, characterized in that: The steps include: S1. Use non-invasive biosensors to collect the user's breathing rate, heart rate and oxygen saturation in real time during sleep to build a physiological data set; S2, filtering, denoising and cleaning the collected original physiological data set to remove noise and outliers; S3, extracting key physiological features from the preprocessed physiological data set, wherein the physiological features include respiratory rate, heart rate variability, and oxygen saturation fluctuation; S4. Using improved deep learning to train a determination model, wherein the determination model performs a comprehensive analysis based on the extracted physiological features to determine whether sleep apnea occurs; S5. The system monitors the user's sleep status in real time and provides immediate feedback or alarm when sleep apnea is detected; S6. The system analyzes the physiological characteristics of different users and provides individual judgment results and health recommendations.

2. The method for determining obstructive sleep apnea based on artificial intelligence according to claim 1, characterized in that: The S1 specifically includes: S11, arranging a plurality of non-invasive biosensors in the user's sleeping environment to collect the user's physiological parameter data during sleep in real time, wherein the physiological parameter data includes the respiratory rate R f , heart rate r and oxygen saturation O s ; S12, the non-invasive biosensor transmits the collected respiratory rate R through the wireless communication module f , heart rate r and oxygen saturation O s The data is transmitted to the data processing terminal, which timestamps each data point and constructs a time series physiological data set D raw : D raw ＝ {(t1,R f (t1),H r (t1),O s (t1)),(t2,R f (t2),H r (t2),O s (t2)),...,(t n ,R f (t n ),H r (t n ),O s (t n ))}; Among them, t i represents the i-th time point, R f (t i ), H r (t i ), O s (t i ) represent the time point t i Respiratory rate, heart rate and oxygen saturation were collected.

3. The method for determining obstructive sleep apnea based on artificial intelligence according to claim 2, characterized in that: The S2 specifically includes: S21. Constructing the time series physiological data set D raw The respiratory rate R f , heart rate r and oxygen saturation O s The data was filtered and a bandpass filter was used to remove high-frequency and low-frequency noise. The filtered time series physiological data set D filtered ; S22, for the filtered time series physiological data set D filtered The data set was denoised and multi-scale analysis was performed on the data using wavelet transform to remove background noise and instantaneous interference. The denoised time series physiological data set D denoised ; S23. Denoised time series physiological data set D denoised The dataset was cleaned to remove outliers and incomplete data, and statistical methods were used to identify and replace abnormal data points. The cleaned time series physiological dataset D clean .

4. The sleep apnea determination method based on artificial intelligence according to claim 3, characterized in that: The S3 specifically includes: S31. From the time series physiological dataset D clean Extract comprehensive physiological features from the respiratory rate R f , heart rate variability HRV and oxygen saturation O s Perform multi-parameter fusion to generate a preliminary comprehensive feature vector, calculate the comprehensive physiological feature vector at each time point in the preliminary comprehensive feature vector, including the time domain and frequency domain characteristics of respiratory frequency, multi-scale entropy of heart rate variability, and high-order statistics of oxygen saturation fluctuations, and generate the final comprehensive feature vector F combined ; S32, combined with multi-parameter fusion technology, use principal component analysis to analyze the comprehensive physiological feature vector F combined Perform dimensionality reduction to obtain the main characteristic component F PCA , the main feature component F after dimensionality reduction PCA As input features, the final feature set for sleep apnea determination is constructed.

5. The method for determining obstructive sleep apnea based on artificial intelligence according to claim 4, characterized in that: The S31 specifically includes: S311. From the time series physiological data set D clean Extract the respiratory rate R f , heart rate variability HRV and oxygen saturation O s Nonlinear features of the proposed algorithm are used to construct a preliminary feature set. S312, respiratory rate R f The radial basis function kernel method is used to capture nonlinear features and obtain the respiratory frequency feature F Rf ; Among them, α j is the weight of the kernel method, σ is the parameter of the basis function, and N is the number of samples; S313, extracting the heart rate variability HRV by kernel method, using polynomial kernel function to capture nonlinear features, and obtaining the heart rate variability feature F HRV : Where c is a constant and d is the degree of the polynomial kernel; S314. Perform kernel method feature extraction on oxygen saturation fluctuation Os, and use Sigmoid kernel function to capture nonlinear features to obtain oxygen saturation feature F Os : F Os =tanh(β(O s (t i ).O s (t j ))+c); Among them, β and c are the parameters of the Sigmoid kernel; S315, using improved variational autoencoder to process and fuse respiratory frequency features F Rf 、Heart rate variability characteristics F HRV and oxygen saturation characteristics F Os , generate the comprehensive feature vector F combined , the loss function of VAE combines sleep breathing data, including reconstruction loss and regularization terms of physiological data features, and is defined as: Among them, x represents the input physiological data feature, z is the latent variable, φ and θ are the parameters of VAE, KL represents the Kullback-Leibler divergence, T i is the time series length of the i-th sample, They represent the reconstructed respiratory rate, heart rate variability and oxygen saturation respectively, λ1 and λ2 are the weights of the regularization term; S316, the final comprehensive feature vector F generated combined : F combined =[F Rf ,F HRV ,F Os ]。 6. The method for determining obstructive sleep apnea based on artificial intelligence according to claim 5, characterized in that: The S32 specifically includes: S321, the comprehensive physiological feature vector F combined All data are standardized, the mean and standard deviation of each feature are calculated, and the feature vector is normalized to zero mean and unit variance. standard ; S322, the standardized feature vector F standard Perform principal component analysis, calculate the characteristic covariance matrix Σ, and solve the eigenvalues ​​and eigenvectors; Among them, μ represents the mean of the standardized feature, and n is the number of samples; S323, select the first k main characteristic components, sort them according to the size of the eigenvalues, select the first k eigenvectors, and construct the main characteristic component matrix W; In=[in1,in2,...,in k ]; Among them, w i represents the i-th main eigenvector, k is the number of main characteristic components selected; S324, the standardized feature vector F standard Projected onto the main characteristic component matrix W, the main characteristic component F after dimensionality reduction is obtained PCA ; F PCA (t i )=F standard (t i )·W; S325, the main feature component F after dimensionality reduction PCA As input features, the final feature set for obstructive sleep apnea determination is constructed.

7. The method for determining obstructive sleep apnea based on artificial intelligence according to claim 6, characterized in that: The S4 specifically includes: S41, using the preprocessed comprehensive feature vector F combined As input features, a deep learning model combining multimodal fusion and graph neural network is constructed for comprehensive analysis and determination of sleep apnea; S42, through multimodal fusion technology, the comprehensive feature vector F combined Processing, integrating the spatial and temporal information of different physiological signals, generating a multimodal feature representation F multi ; F multi =Fusion(F Rf ,F HRV ,F Os ); Among them, Fusion represents the multimodal fusion function, combined with the respiratory rate R f , heart rate variability HRV and oxygen saturation O s feature; S43, multimodal feature representation F multi It is constructed as a heterogeneous graph, where nodes represent physiological characteristics at different time points, edges represent the correlation of time series, and different weights are assigned according to sleep breathing patterns; S44. Use graph neural networks to process heterogeneous graphs, extract complex graph structure features, and generate high-dimensional feature representations F GNN ; F GNN =GNN(F multi ); S45. Pre-train GNN through self-supervised learning technology to construct the obstructive sleep apnea judgment task to maximize the model's ability to capture potential patterns; Among them, y i is the actual label, the presence or absence of sleep apnea, p(y i ) is the predicted probability; S46. Combine contrastive learning strategies to pair samples with similar features and optimize the judgment accuracy of the model; Among them, sim represents the similarity function, which is used to calculate the high-dimensional feature representation F GNN The similarity between them, τ is the temperature parameter, which is used to adjust the sensitivity of contrastive learning; S47, after the training is completed, the optimized deep learning model is used to determine sleep apnea in real time, and the input comprehensive feature vector F combined Pass it to the model to generate the final judgment result; S48. During the model determination process, a final determination result is generated: R final =Decoder(F GNN ); Among them, Decoder is the decoder part of the model, R final This is the final determination result, indicating whether the user has sleep apnea.

8. The sleep apnea determination method based on artificial intelligence according to claim 7, characterized in that: The S48 specifically includes: S481. The high-dimensional feature representation F generated from the graph neural network GNN Input into the decoder model, which includes a multi-layer fully connected neural network; S482, in the first layer of the decoder, the high-dimensional feature representation F of the input GNN Perform linear transformation to generate intermediate feature representation F hidden ; S483, in the second layer of the decoder, the intermediate feature representation H hidden Perform nonlinear transformation to generate output feature representation F output ; S484, in the output layer of the decoder, the output feature representation F output Perform classification operations to generate the final judgment result R final ; S485, in the process of determining the model, combining the threshold setting to determine whether the user has sleep apnea, when R final When the probability value in is greater than the preset threshold, it is determined that sleep apnea occurs; S486, the final judgment result R final And the corresponding probability values ​​are stored in the database.

9. The method for determining obstructive sleep apnea based on artificial intelligence according to claim 7, characterized in that: The S6 specifically includes: S61: construct a user feature database based on the user's historical physiological data and current physiological data, wherein the user feature includes a respiratory rate R f , heart rate variability HRV and oxygen saturation O s ; in, HRV i , They represent the respiratory rate, heart rate variability and oxygen saturation characteristics of the i-th user, respectively, and N is the number of users; S62, using a clustering algorithm to perform cluster analysis on the data in the user feature database, identifying feature patterns of different user groups, and generating a user feature clustering model; Among them, γ ik is an indicator variable. When the i-th user belongs to the k-th cluster, γ ik =1, otherwise γ ik =0;μ k is the center of the kth cluster; S63, matching the physiological characteristics of the current user with the user characteristic clustering model to determine the characteristic group to which the current user belongs; in, HRV current and Respectively represent the current user's breathing rate, heart rate variability and oxygen saturation characteristics; S64, generating a determination result and health advice based on the matching result and the specific physiological characteristics of the current user, the determination result including whether obstructive sleep apnea occurs, and the health advice including specific measures to improve sleep quality; S65, using a recommendation algorithm to provide health advice to the user based on the user's physiological characteristics and the determination result; S66, feeding back the determination results and health advice to the user in real time through the user terminal device, providing guidance on health management and sleep quality improvement for the user; S67. With the user's consent, the user's determination results and health recommendations are stored in a user feature database, and the determination model and recommendation algorithm are optimized.

Images