Detection system for obstructive sleep apnea syndrome

Abstract

The invention relates to the technical field of sleep detection, and discloses a detection system for obstructive sleep apnea syndrome. The system comprises a signal acquisition and enhancement module, a signal segmentation and quality evaluation module, a feature extraction and matching module, a graph structure analysis module and a risk index generation and early warning module. The signal acquisition and enhancement module acquires and enhances a multi-source physiological signal to generate an enhanced signal sequence; the signal segmentation and quality evaluation module adaptively segments the signal to screen out high-quality signal segments; a feature extraction and matching module extracts multi-dimensional feature vectors from the feature extraction and matching module, and the multi-dimensional feature vectors are matched with a standard feature database to obtain a feature difference matrix; the graph structure analysis module deduces the feature node association degree; and the risk module integrates the data, generates a risk index and decides early warning output. The system accurately captures disease characteristics, reduces missed diagnosis and misdiagnosis rates, is suitable for home and large-scale screening, and provides reliable reference for clinical diagnosis.

Description

Technical Field

[0001] This invention relates to the field of sleep detection technology, specifically to a detection system for obstructive sleep apnea syndrome. Background Technology

[0002] Obstructive sleep apnea syndrome (OAS) is a common sleep-disordered breathing condition characterized by recurrent upper airway soft tissue collapse during sleep, leading to apnea or hypoventilation, which in turn causes sleep structure disturbances and decreased blood oxygen saturation. Long-term OAS can induce various chronic diseases such as hypertension, coronary heart disease, and diabetes, and in severe cases, even increase the risk of sudden nocturnal death, posing a significant threat to patients' health and quality of life. With the accelerated pace of modern life, the rising proportion of obese individuals, and the deepening of population aging, the incidence of this condition is showing a year-on-year increasing trend, becoming a public health problem that needs to be addressed.

[0003] Currently, the main method for detecting obstructive sleep apnea syndrome relies on polysomnography (PSG). This method requires a specialized medical facility where multiple sensors are worn by the patient to collect various physiological signals, such as electroencephalograms (EEGs), electrocardiograms (ECGs), blood oxygen saturation, and respiratory airflow. These signals are then analyzed and diagnosed by professionals. However, this traditional method has several limitations: the equipment is bulky and complex to operate, and it requires a stringent testing environment. Patients must complete the monitoring in an unfamiliar sleep environment, which can lead to the "first-night effect" and result in inaccurate results. Furthermore, the testing is expensive and time-consuming, making it difficult to meet the needs of large-scale population screening and long-term home monitoring.

[0004] To address the aforementioned issues, some portable testing devices have emerged on the market, but their performance is often unsatisfactory. Most portable devices can only collect a single type of physiological signal, such as blood oxygen saturation or respiratory airflow signals, failing to comprehensively reflect the patient's physiological state, leading to high rates of missed and false diagnoses. Furthermore, during signal processing, these devices lack effective signal enhancement mechanisms, making it difficult to cope with the weakness of human physiological signals and the influence of external environmental interference, resulting in inconsistent signal quality. In the feature analysis stage, existing technologies mostly employ simple feature extraction and threshold judgment methods, ignoring the intrinsic correlations between physiological signal features and failing to delve into the pathological information hidden behind the features, significantly reducing the reliability and accuracy of the test results. Therefore, developing a detection system for obstructive sleep apnea syndrome capable of accurate acquisition, efficient processing, and in-depth analysis of multi-source signals has become a key problem that needs to be solved by those skilled in the art. Summary of the Invention

[0005] The purpose of this invention is to provide a detection system for obstructive sleep apnea syndrome to solve the problems mentioned in the background art.

[0006] To achieve the above objectives, the present invention provides a detection system for obstructive sleep apnea syndrome, the system comprising: The signal acquisition and enhancement module is used to acquire multi-source physiological signals from the patient and perform signal enhancement operations to generate enhanced signal sequences. The signal segmentation and quality assessment module is used to perform adaptive segmentation processing on the enhanced signal sequence, evaluate the quality score of each signal segment, and select high-quality signal segments based on the quality score. The feature extraction and matching module is used to extract multi-dimensional feature vectors from the high-quality signal segment and perform matching calculations with a standard feature database to obtain a feature difference matrix. The graph structure analysis module is used to analyze the relationships between feature nodes in the feature difference matrix using a graph structure model, and to derive the node correlation degree. The risk index generation and early warning module is used to integrate node correlation and feature difference values ​​to generate a risk index, and to determine the early warning output based on the risk index threshold.

[0007] Preferably, the step of acquiring the patient's multi-source physiological signals and performing signal enhancement specifically includes: acquiring the patient's respiratory waveform signal, blood oxygen concentration signal, and electrocardiogram signal through sensors to form a raw signal set; applying a wavelet transform denoising algorithm to the raw signal set to remove high-frequency noise and baseline drift; adjusting the signal scale using a signal amplitude normalization method to ensure consistent signal amplitude; and smoothing the signal based on sliding window technology to generate an enhanced signal sequence.

[0008] Preferably, the adaptive segmentation of the enhanced signal sequence and the evaluation of the quality score of each signal segment include: dynamically dividing the signal into segments according to the signal frequency characteristics, with the length of each segment being adaptively adjusted based on signal stability; calculating the signal-to-noise ratio and variance of each signal segment as quality evaluation parameters; substituting the quality evaluation parameters into a predefined quality scoring function to output the quality score; setting a quality score threshold and filtering high-quality signal segments with quality scores higher than the threshold.

[0009] Preferably, the step of extracting multi-dimensional feature vectors from the high-quality signal segments includes: performing time-domain analysis on each high-quality signal segment to extract mean, variance, and peak features; performing frequency-domain analysis to extract power spectral density and dominant frequency component features; combining time-domain and frequency-domain features to form a multi-dimensional feature vector; and performing dimensionality reduction processing on the multi-dimensional feature vector to retain the main feature components.

[0010] Preferably, the step of matching the multi-dimensional feature vectors with the standard feature database to calculate the feature difference matrix specifically includes: accessing the standard feature database to obtain the standard feature vectors of healthy individuals; calculating the Euclidean distance and cosine similarity between the multi-dimensional feature vectors and the standard feature vectors; and constructing the feature difference matrix based on the distance and similarity, where the matrix elements represent the difference value of each feature dimension.

[0011] Preferably, the step of using a graph structure model to analyze the relationships between feature nodes in the feature difference matrix and derive the node correlation degree includes: converting the feature difference matrix into a graph structure, where each feature dimension is a node, and the edge weights between nodes are calculated based on the correlation of the difference values; applying a community detection algorithm to identify node communities and analyzing the connection strength of nodes within the community; and calculating the centrality index of each node as the node correlation degree.

[0012] Preferably, the process of generating a risk index by integrating node correlation and feature difference values ​​includes: performing a weighted summation of node correlation, with the weights dynamically adjusted based on feature importance; extracting the maximum and average difference values ​​from the feature difference matrix; and fusing the weighted node correlation with the maximum and average difference values ​​to generate a risk index through a linear combination.

[0013] Preferably, the step of determining the early warning output based on the risk index threshold includes: setting a dynamic risk index threshold, which is adaptively updated according to the patient's historical data; comparing the risk index with the threshold, and triggering an early warning signal when the risk index exceeds the threshold; and generating an early warning report, including the risk index value and the triggering reason.

[0014] Preferably, the dynamic division of signal segments based on signal frequency characteristics includes: The instantaneous frequency of the signal is calculated in real time. When the rate of change of the instantaneous frequency exceeds a preset threshold, it is determined to be a signal characteristic abrupt change point. The signal sequence is divided into multiple signal segments using adjacent mutation points as boundaries; For signal regions with gradual frequency changes, a fixed-length sliding window is used to segment the signal, and the variance of the signal within the window is calculated. If the variance is lower than the stability threshold, the window is defined as a signal segment.

[0015] Preferably, the dimensionality reduction processing of the multi-dimensional feature vector includes: The covariance matrix of the original eigenvectors is calculated using principal component analysis, and its eigenvalues ​​and eigenvectors are solved. Based on the preset cumulative variance contribution rate threshold, the eigenvectors corresponding to the top K largest eigenvalues ​​are selected as principal components. The original multi-dimensional feature vectors are projected onto the selected principal components to form the dimensionality-reduced feature vectors.

[0016] Compared with the prior art, the beneficial effects of the present invention are: The signal acquisition and enhancement module focuses on the collection and optimization of multi-source physiological signals, breaking through the limitations of traditional single-signal acquisition in detection equipment. It can simultaneously acquire multiple physiological signals reflecting a patient's sleep state, ensuring the comprehensiveness of the detection from the data source perspective. Furthermore, addressing the common weakness and susceptibility to external interference in physiological signals, this module performs specialized signal enhancement operations, effectively filtering out noise interference and highlighting the effective information in the signal. This provides high-quality raw data support for subsequent signal processing and analysis, avoiding detection errors caused by poor signal quality.

[0017] The signal segmentation and quality assessment module employs an adaptive segmentation approach, flexibly adjusting segment lengths based on changes in signal characteristics. Compared to fixed-length segmentation methods, this better reflects the dynamic changes in physiological signals, ensuring that each signal segment accurately reflects the physiological state at a specific time. Furthermore, quality assessment filters out high-quality signal segments while directly discarding lower-quality ones, preventing low-quality signals from negatively impacting subsequent feature extraction and analysis. This data preprocessing stage enhances the reliability of the entire detection process, enabling subsequent feature extraction to be based on more valuable signal data.

[0018] The feature extraction and matching module extracts multi-dimensional feature vectors from high-quality signal segments, overcoming the limitation of single-dimensional feature extraction in existing technologies. It comprehensively captures the physiological and pathological information contained in the signal, including features in multiple dimensions such as the time domain and frequency domain, enabling the feature vectors to more completely and accurately represent the patient's sleep breathing state. The feature difference matrix calculated by matching with a standard feature database clearly presents the feature differences between the tested object and health standards, providing an intuitive and comprehensive comparative basis for subsequent risk assessment, and allowing for the precise quantification of disease-related feature differences.

[0019] The graph structure analysis module introduces a graph structure model to analyze the feature difference matrix. It innovatively treats features as nodes and derives node correlation by mining the relationships between feature nodes. This approach allows for in-depth exploration of the potential intrinsic connections between features, rather than viewing each feature in isolation. In the pathological mechanism of obstructive sleep apnea syndrome, there are complex interactions between various physiological features. This module captures these relationships to achieve in-depth mining of pathological information, making the analysis of the disease more closely aligned with its actual pathogenesis and avoiding diagnostic bias caused by ignoring feature correlations.

[0020] The risk index generation and early warning module integrates node correlation and feature difference values ​​to generate a risk index, achieving a comprehensive consideration of multiple assessment indicators. Compared with single-indicator assessments, this is more scientific and comprehensive, and can more accurately reflect the patient's disease risk level. Early warning output is determined based on the risk index threshold, making the early warning mechanism more targeted and reliable. It can promptly report potential disease risks to medical staff and patients, creating favorable conditions for early detection and intervention. It is particularly suitable for large-scale population screening and home monitoring of patients, helping to improve the detection rate of diseases, reduce missed diagnoses and misdiagnoses, and provide strong support for patients' health protection. Attached Figure Description

[0021] Figure 1 This is a schematic diagram illustrating the working principle of the obstructive sleep apnea syndrome detection system described in this invention. Figure 2 This is a flowchart of the signal acquisition and enhancement module; Figure 3 This is a flowchart of feature extraction and dimensionality reduction. Figure 4 A heatmap of multi-dimensional feature difference matrix; Figure 5 This is a trend chart of the obstructive sleep apnea risk index and dynamic threshold. Detailed Implementation

[0022] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0023] Please see Figure 1 This invention provides a detection system for obstructive sleep apnea syndrome. The system includes: a signal acquisition and enhancement module that acquires multi-source physiological signals and performs denoising and smoothing processing to generate an enhanced signal sequence; a signal segmentation and quality assessment module that adaptively segments the enhanced signal sequence and selects high-quality signal segments based on quality scores; a feature extraction and matching module that extracts multi-dimensional feature vectors from the high-quality signal segments and matches them with a standard feature database to generate a feature difference matrix; a graph structure analysis module that converts the feature difference matrix into a graph structure model, analyzes the relationships between feature nodes, and derives the node correlation degree; and a risk index generation and early warning module that integrates node correlation degree and feature difference values ​​to generate a risk index and determines the early warning output based on a threshold.

[0024] Example 1: See Figure 2In practical implementation, the signal acquisition and enhancement module collects the patient's respiratory waveform signals, blood oxygen concentration signals, and electrocardiogram (ECG) signals through sensors to form a raw signal set. The sensors employ medical-grade multi-channel biosignal acquisition equipment. The respiratory waveform signal is acquired through a chest and abdominal respiratory plethysmography sensor, the blood oxygen concentration signal through a fingertip pulse oximeter, and the ECG signal through standard three-lead ECG electrodes. All sensors synchronously acquire data at a sampling frequency of no less than 100Hz to ensure the synchronization and integrity of the time series. In practical implementation, a wavelet transform denoising algorithm is applied to the raw signal set. The wavelet transform denoising algorithm selects the Db4 wavelet as the basis function and performs a 5-level wavelet decomposition on the signal of each channel. Soft thresholding is used to shrink the detail coefficients, effectively removing high-frequency noise and baseline drift. The threshold of the wavelet transform denoising algorithm is adaptively adjusted according to the signal noise level. In practical implementation, a signal amplitude normalization method is used to adjust the signal scale. This method calculates the maximum and minimum values ​​of each signal channel, linearly mapping the signal to the standard range of [-1, 1], thus eliminating amplitude inconsistencies caused by differences in sensor sensitivity or individual physiological differences. Furthermore, a sliding window technique is used to smooth the signal. This technique employs a Hanning window function, with the window length set as an integer multiple of the period corresponding to the signal's dominant frequency. A weighted average is calculated for the data points within the window to generate an enhanced signal sequence.

[0025] In some embodiments, the wavelet transform denoising algorithm includes three steps: signal decomposition, thresholding, and signal reconstruction. The signal decomposition stage decomposes the original signal into approximation coefficients and detail coefficients. The thresholding stage applies an adaptive threshold function to the detail coefficients. The signal reconstruction stage uses the processed coefficients to reconstruct the signal. The threshold selection in the wavelet transform denoising algorithm is based on the principle of unbiased risk estimation, setting different threshold strategies for different physiological signal characteristics. A global threshold is used for respiratory waveform signals, while a hierarchical threshold is used for electrocardiogram signals. In specific implementations, the signal amplitude normalization method includes signal extremum detection and linear transformation. Extremum detection uses a sliding window extremum search algorithm to avoid the influence of instantaneous interference values. The linear transformation formula is that the normalized signal value equals the original signal value minus the minimum value, divided by the difference between the maximum and minimum values, ultimately scaling to the target range.

[0026] It is understandable that implementing the sliding window technique requires consideration of the parameter settings for window length and sliding step size. The window length is dynamically adjusted based on the main frequency components of the signal; a shorter window length is used for signals with complex frequency components, while a longer window length is used for stationary signals. In the weighted average calculation of the sliding window technique, the coefficient distribution of the Hanning window function is a cosine curve with a high center and low sides, effectively reducing spectral leakage. In practical implementation, the generation of the enhanced signal sequence needs to undergo a quality verification process. By calculating the signal-to-noise ratio and waveform integrity indicators, it is ensured that the enhanced signal meets the requirements of subsequent processing.

[0027] Optionally, the acquisition of multi-source physiological signals can employ a distributed sensor network architecture. Each sensor transmits data to a central processing unit (CPU) via wired or wireless means. The CPU is equipped with signal conditioning circuitry, including amplifiers and filters, to perform preliminary conditioning of the raw signals. In specific implementations, the signal acquisition and enhancement module also includes a signal synchronization mechanism. Through hardware timestamps and software interpolation algorithms, strict synchronization between multi-channel signals is ensured, with time synchronization accuracy controlled at the millisecond level. The wavelet transform denoising algorithm can be implemented using a fast wavelet transform algorithm, employing a filter bank to achieve multi-resolution signal analysis and improve computational efficiency.

[0028] In some embodiments, the implementation of the signal amplitude normalization method can be optimized in conjunction with signal characteristics. For non-stationary signals, a segmented normalization strategy is adopted, dividing the signal into multiple segments for separate normalization processing to avoid local feature distortion caused by global normalization. In specific implementations, the parameter optimization of the sliding window technique can be determined experimentally. Parameter lookup tables are established for different types of physiological signals, and the optimal window parameters are automatically selected based on signal characteristics. The enhanced signal sequence is stored using a circular buffer structure to achieve real-time pipelined data processing. The real-time requirements of the signal acquisition and enhancement modules necessitate optimization of algorithm complexity. The computational complexity of the wavelet transform denoising algorithm is reduced by improving the symmetry and compact support properties of the wavelet basis, and the computational load of the signal amplitude normalization method is reduced through incremental computation. In specific implementations, the hardware implementation of the module can employ an embedded digital signal processor, utilizing a parallel computing architecture to accelerate the signal processing flow and ensure that the system can process multi-channel physiological signal data in real time.

[0029] Optionally, the sensor data quality control mechanism includes signal quality index calculation, real-time monitoring of the signal quality of each channel, and automatic calibration or alarm prompts for signal channels that fail to meet quality standards. In practical implementation, the configurability of the signal enhancement process parameters is crucial. Through configuration files or a human-machine interface, the parameters of the wavelet transform denoising algorithm, the signal amplitude normalization method, and the sliding window technique can be adjusted to adapt to the physiological differences of different patients. The output format of the enhanced signal sequence adopts a standardized data packet structure, containing signal data, timestamps, and quality identifiers, facilitating processing and analysis by subsequent modules.

[0030] Example 2: The implementation of the signal segmentation and quality assessment module includes adaptive segmentation of the enhanced signal sequence and evaluation of the quality score of each signal segment. In specific implementation, the adaptive segmentation dynamically divides the signal into segments based on the signal frequency characteristics and calculates the instantaneous frequency of the signal in real time. The instantaneous frequency is calculated using the Hilbert transform method, obtaining the instantaneous frequency value by calculating the phase derivative of the analytical signal. When the rate of change of the instantaneous frequency exceeds a preset threshold, the threshold is set to different values ​​according to the signal type. The threshold is set to 0.5 Hz / s for respiratory waveform signals, 2 Hz / s for electrocardiogram signals, and 0.2 Hz / s for blood oxygen concentration signals. When the rate of change of the instantaneous frequency exceeds the preset threshold, it is determined as a signal feature mutation point. In specific implementation, the signal sequence is divided into multiple signal segments with adjacent mutation points as boundaries. The mutation point detection uses a sliding window difference algorithm with a window length of 100 sampling points. The mutation location is identified by calculating the first-order difference of the instantaneous frequency within the window, and the region between adjacent mutation points is considered as a complete signal segment. For signal regions with gradual frequency changes, a fixed-length sliding window is used for segmentation. The fixed-length sliding window is set to a duration of 2 seconds, corresponding to 200 sampling points. The variance of the signal within the window is calculated using an unbiased estimation method. If the variance is lower than the stability threshold, which is set to one-tenth of the overall signal variance, then the window is defined as a signal segment.

[0031] In practice, the signal-to-noise ratio (SNR) and variance of each signal segment are calculated as quality assessment parameters. The SNR is calculated using the power ratio method, separating the signal and noise bands through bandpass filtering, calculating the ratio of the power values ​​of the two bands and converting it to decibels. The variance is used to calculate the dispersion of all sampling points within a signal segment relative to the mean. These quality assessment parameters are then substituted into a predefined quality scoring function, which is a linear weighted model. The weighting coefficient for the SNR is set to 0.6, and the weighting coefficient for the variance is set to 0.4. The output range of the quality scoring function is normalized to between 0 and 1. A quality score threshold of 0.7 is set to filter high-quality signal segments with scores above the threshold, while signal segments with scores below the threshold are marked as low-quality segments and excluded from further processing.

[0032] In some embodiments, the instantaneous frequency can be calculated using the Short-Time Fourier Transform (SFT) method as an alternative. The window length of the SFT is consistent with the signal segmentation window length. The dominant frequency is determined by spectral peak detection, and then the rate of change of frequency over time is calculated. In specific implementations, the detection of signal feature abrupt changes can be combined with multi-feature judgment. In addition to the instantaneous frequency change rate, the signal amplitude change rate and waveform complexity index can also be introduced to comprehensively determine the location of significant changes in signal features. The segmentation method with a fixed-length sliding window can employ an overlapping window strategy, setting a 50% overlap rate between adjacent windows to prevent effective signal features from being truncated by the window boundaries.

[0033] It's understandable that calculating the variance of a signal segment requires excluding the influence of outliers. Before calculating the variance, outlier detection should be performed on the signal segment, using the three-standard-deviation principle to identify and remove sampling points that significantly deviate from the normal range. In practice, the parameters of the quality scoring function can be optimized using machine learning methods. A logistic regression model can be trained using signal segment samples of known quality, automatically learning the optimal combination of weighting coefficients for the signal-to-noise ratio and variance. The quality score threshold can be set considering individual differences, dynamically adjusting the threshold level based on the patient's historical signal quality data to improve the accuracy of screening.

[0034] Optionally, adaptive segmentation processing can incorporate multi-resolution analysis, simultaneously segmenting the signal at different scales. Coarse-scale segmentation captures macroscopic feature changes, while fine-scale segmentation captures microscopic feature fluctuations. In practice, signal segment boundary processing employs a smooth transition technique, setting a 10-millisecond transition region at the segment boundaries and using a cosine smoothing function to avoid signal discontinuities at the boundaries. The quality assessment process can include waveform morphology consistency checks and calculating the correlation coefficient between the signal segment and the standard template as auxiliary quality assessment parameters.

[0035] In some embodiments, signal-to-noise ratio (SNR) calculation can employ wavelet transform, transforming the signal to different scale spaces through wavelet decomposition. Noise power is estimated in detail coefficients, and signal power is estimated in approximation coefficients. In specific implementations, variance index calculation can utilize an improved robust variance estimation method, using the median absolute deviation instead of traditional variance calculation to improve tolerance to outliers. The application of quality score thresholds can employ a soft threshold strategy, setting two threshold levels. Signal segments with scores above the upper threshold are directly accepted, those below the lower threshold are directly rejected, and those between the two thresholds enter a manual review process. The real-time requirements of signal segmentation necessitate optimized computational processes. Instantaneous frequency calculation uses a recursive filter to implement Hilbert transform, reducing computational complexity. In specific implementations, quality assessment parameters can be calculated using a parallel processing architecture, simultaneously calculating the SNR and variance indexes of multiple signal segments, leveraging multi-core processors to accelerate processing. The screening results of high-quality signal segments are stored as a timestamped index table, facilitating rapid location and retrieval of corresponding signal data by subsequent modules.

[0036] Optionally, the signal segment length can be constrained with upper and lower limits: the minimum segment length is no less than 0.5 seconds, and the maximum segment length is no more than 10 seconds, to avoid excessively fragmented or lengthy signal segments. In practice, the quality scoring function can incorporate a clinical validity check of the signal segments, using a rule engine to determine whether the signal segment contains clinically significant physiological events, such as a complete respiratory cycle or cardiac cycle. The parameters for adaptive segmentation processing can be dynamically configured according to the signal type; a larger segment length is used for stable blood oxygen concentration signals, while a smaller segment length is used for rapidly changing respiratory waveform signals.

[0037] In practical implementation, the feedback mechanism for signal segment quality assessment results is crucial. Characteristic information of low-quality signal segments is recorded and the causes analyzed, such as sensor detachment and motion artifacts, to improve the signal acquisition process. The output of the signal segmentation and quality assessment module adopts a standard data structure, with each signal segment accompanied by complete metadata information, including start time, duration, quality score, and original values ​​of assessment parameters. During module implementation, a signal segment quality database is established to accumulate signal quality data from different patients under various physiological states, providing a data foundation for adaptive optimization of quality assessment parameters.

[0038] Example 3: See Figure 3In practical implementation, time-domain analysis is performed on each high-quality signal segment to extract mean, variance, and peak features. The mean feature is obtained by calculating the arithmetic mean of all sampling points within the signal segment. The variance feature is obtained by calculating the average of the squared deviations of the sampling points from the mean within the signal segment. The peak feature is obtained by identifying local maxima and minima within the signal segment and calculating their amplitude differences. A sliding window method is used in the time-domain analysis to ensure the continuity of feature extraction, with the window size consistent with the signal segment length. In practical implementation, frequency-domain analysis is performed to extract power spectral density and dominant frequency component features. The power spectral density feature is calculated by converting the signal segment to the frequency domain using a fast Fourier transform and then calculating the power value at each frequency point. The dominant frequency component feature is obtained by detecting the peak frequency in the power spectrum. A Hanning window is used in the frequency-domain analysis to reduce spectral leakage, and the frequency resolution is adaptively adjusted according to the signal segment length. In practice, time-domain and frequency-domain features are combined to form a multi-dimensional feature vector. The time-domain features include mean features, variance features, and peak features, while the frequency-domain features include power spectral density features and dominant frequency component features. The dimension of the multi-dimensional feature vector is equal to the sum of the number of time-domain features and the number of frequency-domain features. The feature values ​​are standardized to eliminate the influence of dimensions.

[0039] In some embodiments, the mean feature calculation for time-domain analysis can employ a weighted averaging method, assigning different weights to sampling points within a signal segment based on their temporal location, with higher weights for center points and lower weights for edge points. In specific implementations, variance feature calculation can use an unbiased estimation method, with the denominator being the number of sampling points minus one to improve estimation accuracy. Peak feature identification employs a derivative detection algorithm, locating extreme points by calculating the zero-crossing points of the first derivative of the signal. Power spectral density estimation for frequency-domain analysis can use the Welch method, dividing the signal segment into overlapping sub-segments for averaging to reduce variance. Major frequency component feature extraction can include multiple peak frequencies, such as the top three main peak frequencies, to capture richer frequency domain information. During joint feature formation, time-domain and frequency-domain features are concatenated in a preset order, for example, arranging time-domain features first followed by frequency-domain features, forming a unified multi-dimensional feature vector.

[0040] It is understandable that the standardization of multi-dimensional feature vectors is crucial. The Z-score standardization method is used to convert each feature value to zero mean and unit variance. The standardization parameters are learned from the training data and applied to new data. In specific implementation, dimensionality reduction is performed on the multi-dimensional feature vectors. Principal component analysis (PCA) is used to calculate the covariance matrix of the original feature vectors. The covariance matrix reflects the linear relationship between features. The eigenvalues ​​and eigenvectors of the covariance matrix are solved. The eigenvalues ​​represent the magnitude of the variance contribution of the principal components, and the eigenvectors represent the direction of the principal components. Based on a preset cumulative variance contribution rate threshold, the eigenvectors corresponding to the K largest eigenvalues ​​are selected as principal components. The cumulative variance contribution rate threshold is set to 85%, and the value of K is determined by calculating the cumulative sum of the eigenvalues ​​when the threshold is reached. The original multi-dimensional feature vectors are projected onto the selected principal components to form the dimensionality-reduced feature vectors. The projection process is implemented through matrix multiplication.

[0041] In some embodiments, the implementation of the principal component analysis method includes a data centering step, which first subtracts the mean vector from the original eigenvectors to ensure zero-centering of the data. The formula for calculating the covariance matrix is: in: Represents the covariance matrix. Indicates the number of eigenvectors. Represents the centered feature matrix, with the symbol... This represents the matrix transpose operation. Eigenvalues ​​and eigenvectors are solved using the Jacobi iteration method or the QR algorithm to ensure numerical stability. The projection operation uses the following formula: in: This represents the eigenvectors after dimensionality reduction. Represents the original eigenvector matrix. This represents the principal component matrix composed of the first K eigenvectors.

[0042] Optionally, dimensionality reduction can be combined with other feature selection methods, such as linear discriminant analysis, to further improve the representativeness of the features. In practice, the feature vectors of principal component analysis are sorted in descending order of eigenvalues ​​to ensure that key information is retained first. The dimensionality of the feature vectors after dimensionality reduction is significantly reduced, but most of the original variance is retained, facilitating subsequent matching calculations. The cumulative variance contribution rate threshold can be adjusted according to application requirements; a high threshold retains more information but increases dimensionality, while a low threshold reduces dimensionality but may lose details. The computational efficiency of the feature extraction process needs to be optimized. Time-domain and frequency-domain analysis adopt a parallel computing architecture to process multiple high-quality signal segments simultaneously, and the eigenvalue decomposition of principal component analysis is accelerated using an efficient numerical library. In practice, the storage of multi-dimensional feature vectors adopts a compact data structure, with each feature vector accompanied by metadata such as signal segment identifiers and timestamps. The dimensionality-reduced feature vectors are used for matching with a standard feature database to ensure the overall performance of the detection system. The feature extraction and dimensionality reduction modules integrate a caching mechanism to store intermediate results to reduce redundant calculations, supporting real-time and batch processing modes.

[0043] In some embodiments, the time-domain peak features can be refined to include dynamic characteristics such as rise time and fall time, and the frequency-domain dominant frequency components can be expanded to include spectral characteristics such as bandwidth and harmonic ratio. In specific implementations, dimensionality reduction of multi-dimensional feature vectors can be achieved by introducing nonlinear methods such as kernel principal component analysis to handle complex relationships between features, but linear principal component analysis is preferred for computational simplicity.

[0044] Example 4: In specific implementation, the step of matching multi-dimensional feature vectors with a standard feature database to obtain a feature difference matrix is ​​achieved by accessing the standard feature database through a database interface module. The standard feature database stores a set of standard feature vectors from healthy individuals. This set of standard feature vectors originates from a reference dataset established through large-scale clinical studies. Each standard feature vector contains the same feature dimensions as the patient's multi-dimensional feature vectors, arranged in a preset order of time-domain and frequency-domain features. The matching calculation process first performs standardization preprocessing on the patient's multi-dimensional feature vectors to ensure they are on the same numerical scale as the vectors in the standard feature database. Then, the Euclidean distance between the multi-dimensional feature vectors and each standard feature vector is calculated using the summation and square root of squared differences. Simultaneously, the cosine similarity between the multi-dimensional feature vectors and each standard feature vector is calculated using the vector dot product divided by the product of their magnitudes. A feature difference matrix is ​​constructed based on Euclidean distance and cosine similarity. The rows of the feature difference matrix correspond to the multi-dimensional feature vectors of patients, and the columns of the feature difference matrix correspond to different feature dimensions in the standard feature database. The matrix elements of the feature difference matrix are composed of weighted combinations of Euclidean distance and cosine similarity, and the weight coefficients are assigned according to the importance of the feature dimensions.

[0045] In practice, referring to Table 1, the feature difference matrix is ​​constructed using normalization to ensure the comparability of difference values ​​across different dimensions. The Euclidean distance component is first normalized to the [0,1] interval using a minimum-maximum normalization method, and the cosine similarity component is converted to dissimilarity and then normalized. Each element of the feature difference matrix represents the degree of deviation between the patient's characteristics and the standard characteristics on a specific feature dimension. The feature difference matrix is ​​stored using a two-dimensional array data structure, simultaneously saving feature dimension identification information and metadata for difference value calculation.

[0046] Table 1: Parameters for Calculating the Feature Difference Matrix It is understandable that the computational efficiency of the feature difference matrix is ​​improved through matrix operation optimization. A batch processing approach is used to simultaneously calculate the difference values ​​across multiple feature dimensions, and a linear algebra library is utilized to accelerate matrix multiplication operations. In practical implementation, the standard feature database update mechanism periodically incorporates new healthy population data. A sliding window approach is used to maintain the timeliness and representativeness of the standard feature vectors, and the computation results of the feature difference matrix are cached in a temporary storage area for use by the graph structure analysis module.

[0047] In its implementation, the graph structure analysis module uses a graph structure model to analyze the relationships between feature nodes in the feature difference matrix. The process of deriving node correlation first converts the feature difference matrix into a graph structure. Nodes in the graph structure correspond to each feature dimension of the feature difference matrix. Node attributes include the feature dimension name, feature type, and difference statistics. Edges in the graph structure connect nodes of different feature dimensions. Edge weights are calculated based on the correlation of difference values ​​across different feature dimensions in the feature difference matrix. The correlation calculation uses the Pearson correlation coefficient method to calculate the linear correlation between the sequences of difference values ​​across different feature dimensions. After the graph structure is constructed, a community detection algorithm is applied to identify node clusters. The community detection algorithm uses the Louvain algorithm, based on the principle of maximizing modularity, to iteratively optimize node grouping and analyze the connection strength of nodes within a cluster. The connection strength is obtained by calculating the average of all edge weights within the cluster. Finally, the centrality index of each node is calculated as the node correlation. The centrality index is calculated using the eigenvector centrality algorithm, considering the importance of the node itself and the importance of the nodes it connects to. The node correlation value is normalized to between 0 and 1.

[0048] Optionally, the graph structure model can be represented as a weighted undirected graph, where node size is proportional to node affinity and edge thickness is proportional to edge weight, visually displaying the relationship patterns between feature nodes. In specific implementations, the calculation of edge weights can incorporate nonlinear correlation metrics such as mutual information as a supplement to capture non-monotonic dependencies between feature dimensions. Community detection algorithms can also use spectral clustering as an alternative, partitioning nodes based on the eigenvalue decomposition of the graph Laplacian matrix. The calculation of node affinity can combine multiple centrality indicators such as betweenness centrality and compactness centrality, obtaining a comprehensive node affinity through weighted fusion.

[0049] In some embodiments, the transformation from the feature difference matrix to the graph structure can be achieved by setting a correlation threshold, retaining only edges with an absolute correlation value higher than the threshold, reducing redundant connections in the graph and improving computational efficiency. The correlation threshold is typically set between 0.3 and 0.5. In specific implementations, node community analysis can identify functional modules between feature dimensions, such as time-domain feature communities and frequency-domain feature communities. Intra-community connection strength is calculated using the average edge weight method, while inter-community connection strength is calculated by averaging the boundary edge weights. Node correlation calculation employs an iterative algorithm, initially assigning each node the same weight, and then iteratively updating based on the weights of neighboring nodes until convergence. The computational complexity of graph structure analysis is related to the square of the number of nodes. For high-dimensional features, approximation algorithms or sampling methods are needed to improve computational speed. In specific implementations, the graph structure analysis module uses a distributed graph computing framework to handle large-scale feature dimensions, and the node correlation results are stored in vector form for use by the risk index generation module. The visual output of the graph structure model helps to understand the intrinsic relationships between feature dimensions, providing intuitive support for clinical interpretation.

[0050] Optionally, graph structure analysis can incorporate a dynamic update mechanism, adjusting the graph topology and node relationships as new data flows in. The persistent storage of the graph structure uses a graph database format, preserving complete node and edge attribute information. In practice, prior knowledge weights can be introduced into the calculation of node relationships, assigning higher initial weights to clinically significant feature dimensions, thus influencing the final node relationship calculation results. A feedback mechanism is established between the graph structure analysis module and the feature extraction module to adjust the feature extraction strategy based on the node relationship results, optimizing the overall performance of the detection system.

[0051] See Figure 4This chart, a core output of the feature extraction and matching module in the obstructive sleep apnea syndrome detection system, visually presents the differences between the multi-dimensional characteristics of patients and the standard characteristics of healthy individuals in the form of a heatmap. The horizontal axis represents the standard feature dimension, and the vertical axis represents the patient feature dimension; the color intensity and numerical values ​​together reflect the feature differences. The technical value of this chart lies in its ability to overcome the limitations of traditional single-feature comparisons. Through a matrix-based multi-dimensional difference depiction, it not only visually presents the deviation patterns of patients' pathological characteristics but also provides a refined quantitative basis for the risk index generation module, thus contributing to the accurate detection and mechanism analysis of obstructive sleep apnea syndrome.

[0052] Example 5: In specific implementation, the process of integrating node correlation and feature difference values ​​to generate a risk index begins with a weighted summation operation on node correlation. The node correlation data comes from the normalized correlation vector output by the graph structure analysis module. The weights of the weighted summation are dynamically adjusted based on feature importance, which is obtained through historical data analysis. Feature importance assessment considers the discriminative power and clinical significance of feature dimensions in past cases. The weighted summation calculation uses an element-wise multiplication and summation method. The formula is expressed as: the weighted node correlation equals the sum of the products of each node correlation multiplied by its corresponding weight coefficient. The sum of the weight coefficients is normalized to 1 to ensure consistency in the scale of the results. In specific implementation, the maximum difference value and the average difference value are extracted from the feature difference matrix. The maximum difference value is found by traversing all elements of the feature difference matrix to find the largest difference value. The average difference value is obtained by calculating the arithmetic mean of all elements of the feature difference matrix. The maximum difference value reflects the most significant single-dimensional anomaly, and the average difference value reflects the overall degree of deviation. In practice, the weighted node correlation degree is integrated with the maximum difference value and the average difference value, and a risk index is generated through linear combination. The coefficients of the linear combination are determined by regression analysis, which is trained using a historical dataset with known clinical outcomes. The linear combination formula is: the risk index equals the weighted node correlation degree multiplied by coefficient A, plus the maximum difference value multiplied by coefficient B, plus the average difference value multiplied by coefficient C. The coefficients A, B, and C are fitted to obtain the optimal value using the least squares method.

[0053] It is understandable that the parameter settings in the risk index generation process require clinical validation. Cross-validation is used to determine the weighting coefficients and linear combination coefficients to avoid overfitting, and the numerical range of the risk index is limited to 0 to 100 through scaling for easier interpretation. In practice, the risk index generation module establishes a parameter configuration table to store all weighting coefficients and combination coefficients. This parameter configuration table supports adjustments based on different patient group characteristics, such as setting different parameter combinations for different age groups or genders.

[0054] In practical implementation, the early warning output is determined based on the risk index threshold. This includes setting a dynamic risk index threshold, which is adaptively updated based on the patient's historical data. The adaptive update algorithm uses an exponentially weighted moving average method. The new threshold equals the historical threshold multiplied by a decay factor plus the current risk index multiplied by (1 minus the decay factor). The decay factor is set to 0.9, giving historical data a significant weight to maintain threshold stability. In practical implementation, the risk index is compared with the threshold. When the risk index exceeds the threshold, an early warning signal is triggered. The early warning signal triggering mechanism uses a hysteresis comparator design, setting rising and falling thresholds to avoid frequent switching at threshold boundaries. The rising threshold is 5 percentage points higher than the falling threshold to prevent fluctuations. In practical implementation, an early warning report is generated, including the risk index value and the triggering reason. Triggering reason analysis is achieved by identifying the feature dimension that contributes most to the risk index. The contribution calculation uses the Shapley value allocation method to fairly distribute the contribution ratio of each feature dimension.

[0055] In some embodiments, the updating of the dynamic risk index threshold can take into account the time factor, giving higher weight to recent data to reflect changes in patient status more quickly, and setting the threshold update cycle to 24 hours to avoid overly frequent adjustments. In specific implementations, the generation of early warning reports can integrate natural language processing technology to convert the contribution of feature dimensions into clinical terminology descriptions, such as converting "abnormal contribution of respiratory waveform variance features 35%" into "irregular respiratory rhythm is a major risk factor".

[0056] Optionally, a nonlinear fusion method can be introduced to generate the risk index. When there is an interaction effect between the correlation between weighted nodes and the difference value, a neural network model is used for fusion. The neural network model structure is a three-layer fully connected network. The three nodes in the input layer correspond to the three components, the five nodes in the hidden layer use the ReLU activation function, and the one node in the output layer outputs the risk index. In specific implementation, a multi-level warning mechanism can be set to trigger the warning signal. Different warning levels are divided according to the degree to which the risk index exceeds the threshold. For example, yellow warning, orange warning, and red warning correspond to different clinical response procedures.

[0057] In some embodiments, the initial value of the dynamic risk index threshold can be derived from population baseline data, taking a specific percentile as the initial threshold based on the risk index distribution of healthy individuals, and then gradually transitioning to an individualized threshold. In specific implementations, the output format of the warning report uses a standardized template, including basic patient information, detection timestamp, risk index value, risk level, list of major abnormal features, and clinical recommendations. The warning report supports multiple output methods, including screen display, printout, and electronic transmission.

[0058] Optionally, the risk index generation process can incorporate quality control checks. When input data quality indicators fall below the standard, the reliability of the risk index is flagged as reduced, and data quality notes are added to the warning report. In implementation, the weighted summation of node correlations can employ an adaptive weight adjustment mechanism. When data for certain feature dimensions is missing, the weight allocation is automatically adjusted to ensure the robustness of the risk index calculation. After a warning signal is triggered, the system automatically records the trigger event to the log database, including the trigger time, risk index value, threshold level, and main contributing characteristics, for subsequent analysis and system optimization.

[0059] Optionally, the risk index generation and early warning module can be configured with a manual review process. For high-risk warnings, clinical staff must confirm before automatic issuance. The manual review interface displays detailed feature analysis charts and trend information. In implementation, the early warning output interface is integrated with the hospital information system, supporting the automatic creation of electronic medical record entries and the generation of notification messages for medical staff. Early warning message transmission uses a priority queue mechanism, with high-risk warnings sent first. The risk index calculation process undergoes periodic performance evaluation. The sensitivity and specificity of threshold settings are analyzed through receiver operation characteristic curve analysis, and parameter configuration is optimized based on the evaluation results.

[0060] See Figure 5 This chart is the core visualization output of the risk index generation and early warning module of the obstructive sleep apnea syndrome detection system. It dynamically presents the changes in the patient's disease risk and early warning status during nighttime sleep. The horizontal axis represents the detection time, and the vertical axis represents the risk index. The evolution of risk and the early warning logic are intuitively displayed through three elements: red broken lines, blue dashed lines, and yellow areas. The technical value of this chart lies in its ability to quantify the fluctuation pattern of obstructive sleep apnea risk during nighttime sleep by comparing the time-series risk index with dynamic thresholds. It also intuitively reflects the real-time decision-making process of the early warning module, providing a visual basis for medical staff to quickly assess the patient's risk level and formulate intervention strategies. This effectively solves the pain points of delayed risk assessment and untimely early warning in traditional detection methods.

[0061] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A detection system for obstructive sleep apnea syndrome, characterized in that, The system includes: The signal acquisition and enhancement module is used to acquire multi-source physiological signals from the patient and perform signal enhancement operations to generate enhanced signal sequences. The signal segmentation and quality assessment module is used to perform adaptive segmentation processing on the enhanced signal sequence, evaluate the quality score of each signal segment, and select high-quality signal segments based on the quality score. The feature extraction and matching module is used to extract multi-dimensional feature vectors from the high-quality signal segment and perform matching calculations with a standard feature database to obtain a feature difference matrix. The graph structure analysis module is used to analyze the relationships between feature nodes in the feature difference matrix using a graph structure model, and to derive the node correlation degree. The risk index generation and early warning module is used to integrate node correlation and feature difference values ​​to generate a risk index, and to determine the early warning output based on the risk index threshold.

2. The detection system for obstructive sleep apnea syndrome according to claim 1, characterized in that, The steps of acquiring the patient's multi-source physiological signals and performing signal enhancement operations specifically include: acquiring the patient's respiratory waveform signal, blood oxygen concentration signal, and electrocardiogram signal through sensors to form a raw signal set; applying a wavelet transform denoising algorithm to the raw signal set to remove high-frequency noise and baseline drift; adjusting the signal scale using a signal amplitude normalization method to ensure consistent signal amplitude; and smoothing the signal based on sliding window technology to generate an enhanced signal sequence.

3. The detection system for obstructive sleep apnea syndrome according to claim 2, characterized in that, The adaptive segmentation of the enhanced signal sequence and the evaluation of the quality score of each signal segment include: dynamically dividing the signal into segments based on the signal frequency characteristics, with the length of each segment adaptively adjusted based on signal stability; calculating the signal-to-noise ratio and variance of each signal segment as quality evaluation parameters; substituting the quality evaluation parameters into a predefined quality scoring function to output the quality score; setting a quality score threshold and filtering high-quality signal segments with quality scores higher than the threshold.

4. The detection system for obstructive sleep apnea syndrome according to claim 3, characterized in that, The step of extracting multi-dimensional feature vectors from the high-quality signal segments includes: performing time-domain analysis on each high-quality signal segment to extract mean, variance, and peak features; performing frequency-domain analysis to extract power spectral density and dominant frequency component features; combining time-domain and frequency-domain features to form a multi-dimensional feature vector; and performing dimensionality reduction processing on the multi-dimensional feature vector to retain the main feature components.

5. The obstructive sleep apnea syndrome detection system according to claim 4, characterized in that, The steps for matching multi-dimensional feature vectors with a standard feature database to calculate the feature difference matrix specifically include: accessing the standard feature database to obtain standard feature vectors of healthy individuals; calculating the Euclidean distance and cosine similarity between the multi-dimensional feature vectors and the standard feature vectors; and constructing a feature difference matrix based on the distance and similarity, where each element represents the difference value of each feature dimension.

6. The detection system for obstructive sleep apnea syndrome according to claim 5, characterized in that, The method of using a graph structure model to analyze the relationships between feature nodes in the feature difference matrix and derive node correlation includes: converting the feature difference matrix into a graph structure, where each feature dimension is a node, and the edge weights between nodes are calculated based on the correlation of the difference values; applying a community detection algorithm to identify node communities and analyzing the connection strength of nodes within the community; and calculating the centrality index of each node as the node correlation.

7. The detection system for obstructive sleep apnea syndrome according to claim 6, characterized in that, The process of integrating node correlation and feature difference values ​​to generate a risk index includes: weighting and summing node correlations, with the weights dynamically adjusted based on feature importance; extracting the maximum and average difference values ​​from the feature difference matrix; and fusing the weighted node correlations with the maximum and average difference values ​​to generate a risk index through linear combination.

8. The detection system for obstructive sleep apnea syndrome according to claim 7, characterized in that, The method of determining the early warning output based on the risk index threshold includes: setting a dynamic risk index threshold, which is adaptively updated according to the patient's historical data; comparing the risk index with the threshold, and triggering an early warning signal when the risk index exceeds the threshold; and generating an early warning report, including the risk index value and the triggering reason.

9. The detection system for obstructive sleep apnea syndrome according to claim 3, characterized in that, The dynamic segmentation of signal segments based on signal frequency characteristics includes: The instantaneous frequency of the signal is calculated in real time. When the rate of change of the instantaneous frequency exceeds a preset threshold, it is determined to be a signal characteristic abrupt change point. The signal sequence is divided into multiple signal segments using adjacent mutation points as boundaries; For signal regions with gradual frequency changes, a fixed-length sliding window is used to segment the signal, and the variance of the signal within the window is calculated. If the variance is lower than the stability threshold, the window is defined as a signal segment.

10. The detection system for obstructive sleep apnea syndrome according to any one of claims 4 or 5, characterized in that, The dimensionality reduction process for the multi-dimensional feature vector includes: The covariance matrix of the original eigenvectors is calculated using principal component analysis, and its eigenvalues ​​and eigenvectors are solved. Based on the preset cumulative variance contribution rate threshold, the eigenvectors corresponding to the top K largest eigenvalues ​​are selected as principal components. The original multi-dimensional feature vectors are projected onto the selected principal components to form the dimensionality-reduced feature vectors.

Images