A snore-based real-time apnea and hypopnea prediction method

Abstract

The present application provides a kind of real-time apnea and hypopnea prediction method based on snoring, comprising: obtaining the snoring audio and sleep index data of target object, pre-processing and feature extraction are carried out on the snoring audio, the respiratory event time node marked in polysomnography data is used as a label to extract positive and negative samples, based on sleep index data, positive and negative sample extraction results, predict the probability of occurrence of apnea and hypopnea in subsequent preset time;For obtaining the snoring audio and physical examination data of target object and polysomnography data, based on the feature fusion mode, it is modeled in the unified feature space and combined with the principle of feature importance, the occurrence and nonoccurrence of apnea and hypopnea and the corresponding probability are accurately predicted in real time, monitoring is carried out before the respiratory event occurrence node, and corresponding warning signal is sent according to the detection result, and the target object is actively intervened.

Description

Technical Field

[0001] The present invention relates to the technical field of prediction methods, and in particular to a real-time apnea and hypopnea prediction method based on snoring. Background Art

[0002] Obstructive sleep apnea hypopnea syndrome (OSAHS) refers to recurrent apnea and hypopnea caused by upper airway collapse and obstruction during sleep. Clinical manifestations include snoring, waking up from sleep apnea, disrupted sleep structure, frequent drops in blood oxygen saturation, often accompanied by increased nocturia, morning headaches, daytime sleepiness, memory loss, and, in severe cases, decreased cognitive function and abnormal behavior. OSAHS is a systemic disease. Recurrent nocturnal apnea and hypopnea cause chronic intermittent hypoxia and carbon dioxide retention, leading to metabolic diseases such as hypertension, cardiovascular and cerebrovascular diseases, and type 2 diabetes. It is also a major cause of sudden death and road traffic accidents, making it a serious social problem. According to data, the number of people suffering from OSAHS worldwide has reached 1 billion, and a large number of patients in China remain undiagnosed.

[0003] The standard diagnostic method for OSAHS is polysomnography (PSG). This examination needs to be completed in a specialized sleep laboratory with a sleep technician on duty all night. Usually, the signals recorded by PSG include electroencephalogram, electrooculogram, chin electromyogram, lower limb electromyogram, respiratory airflow signal, respiratory effort signal, blood oxygen saturation, body position, electrocardiogram, etc. Based on the respiratory-related signals, the situation of respiratory events occurring during the patient's sleep can be obtained, among which the apnea-hypopnoea index (AHI) is an important diagnostic basis. Respiratory events specifically include obstructive apnea, central apnea, mixed apnea, hypopnea, and respiratory event-related micro-arousals. Although OSAHS is a common disease in my country, the PSG equipment is complex, the monitoring cost is high, and the examination requirements are strict. Professional sleep technicians are needed to connect cumbersome lines and analyze data from multiple signals throughout the night. This greatly reduces the practicality and universality of PSG equipment. Currently, only tertiary hospitals or some secondary hospitals in large and medium-sized cities can perform standardized diagnosis and treatment of this disease, resulting in a large number of patients not receiving timely diagnosis and treatment, causing great harm to people's health.

[0004] During overnight PSG monitoring, patients must be covered in sensors, significantly reducing their comfort level. Furthermore, this examination is subject to numerous limitations, including monitoring time and accuracy, making long-term observation impossible. Therefore, methods for screening for OSAHS and predicting respiratory events at home using mobile phones and wearable devices are urgently needed. This is driven by rising public health awareness and the gradual improvement of wearable device monitoring technology from various manufacturers, ensuring the future applicability of the technology used in this proposal.

[0005] Although obstructive sleep apnea-hypopnea syndrome is a common disease in my country, it requires professional sleep technicians to connect cumbersome lines and analyze data from multiple signals throughout the night, resulting in low diagnostic efficiency and low diagnostic accuracy. At the same time, current studies rarely consider real-time early warning and intervention of respiratory events.

[0006] Therefore, the present invention provides a real-time apnea and hypopnea prediction method based on snoring, establishes a respiratory event prediction mechanism based on snoring and other real-time sleep indicators, and accurately and in real time issues early warnings for the existence of respiratory events after the decision point. In combination with the principle of feature importance, active intervention methods (such as hypoglossal nerve stimulation and postural intervention) are designed, which will help further explore the changes in the upper airway before a respiratory event in the field of medical and engineering integration and help implement a real-time early warning intervention system. Against the backdrop of increasing attention paid to sleep problems by the public health and society, data support is provided for the subjective judgment of physicians on respiratory changes caused by changes in the upper airway before and after respiratory events, laying a solid algorithmic foundation for the subsequent construction of a complete system. Summary of the Invention

[0007] The present invention provides a real-time apnea and hypopnea prediction method based on snoring, which is used to obtain snoring audio, physical examination data and polysomnography monitoring data of a target subject, model them in a unified feature space based on feature fusion and combine with the principle of feature importance, to accurately predict the occurrence of apnea and hypopnea and their corresponding probabilities in real time, monitor before the occurrence of respiratory events, issue corresponding warning signals based on the detection results, and actively intervene in the target subject.

[0008] The present invention provides a real-time apnea and hypopnea prediction method based on snoring, comprising:

[0009] S1: Obtain the target subject’s snoring audio and sleep index data;

[0010] S2: Preprocessing and feature extraction of the snoring audio;

[0011] S3: Extract positive and negative samples using the respiratory event time nodes annotated in the polysomnography data as annotations;

[0012] S4: Based on sleep index data and positive and negative sample extraction results, predict the probability of apnea and hypopnea within the subsequent preset time.

[0013] Preferably, the real-time apnea and hypopnea prediction method based on snoring, S2: preprocessing and feature extraction of the snoring audio, comprises:

[0014] performing noise reduction processing on the snoring audio based on multi-window spectral subtraction and Wiener filtering, and then performing pre-emphasis processing to obtain a pre-processed snoring audio signal;

[0015] Non-snoring frames and breathing frames are filtered out based on preset hard discrimination and soft discrimination methods.

[0016] Preferably, the real-time apnea and hypopnea prediction method based on snoring, S2: preprocessing and feature extraction of the snoring audio, further comprises:

[0017] Calculating the background signal threshold of the preprocessed snoring audio signal;

[0018] Dividing the preprocessed snoring audio signal into a plurality of audio segments of a preset length, extracting acoustic features of each audio segment, and determining a candidate region for selecting a snoring segment start point and a candidate region for selecting a snoring segment end point based on the acoustic features of each audio segment;

[0019] The frame corresponding to the maximum value of the inverse difference function of the starting point candidate area is used as the starting point of the snoring segment, and the frame corresponding to the maximum value of the inverse difference function of the ending point candidate area is used as the ending point of the snoring segment;

[0020] determining all snoring segments in the continuous snoring audio signal based on adjacent starting points and ending points;

[0021] Obtaining the short-time energy and zero-crossing rate corresponding to all snoring frames, and determining whether the short-time amplitude of the snoring frame is greater than a minimum energy threshold or the zero-crossing rate is greater than a zero-crossing rate threshold; if so, determining the snoring frame as a valid snoring frame; otherwise, determining the snoring frame as an invalid snoring frame;

[0022] Obtaining a time interval between adjacent valid snoring frames in the continuous snoring audio signal, and when the time interval is less than a preset time threshold, merging a first valid snoring frame and a second valid snoring frame in the adjacent valid snoring frames to obtain a merged audio signal;

[0023] Otherwise, do not merge.

[0024] Preferably, the real-time apnea and hypopnea prediction method based on snoring, S2: preprocessing and feature extraction of the snoring audio, further comprises:

[0025] Inputting the combined audio into a deep neural network to extract a Mel-frequency cepstral coefficient feature matrix;

[0026] Determining a mutation point of the merged audio based on a first preset criterion, and dividing the merged audio into a plurality of audio samples based on the mutation point;

[0027] determining an arrangement order of valid snore frames in each audio sample, wherein each audio sample is composed of a preset number of consecutive valid snore frames;

[0028] Clustering valid snoring frames in the same arrangement order in each audio sample based on the Mel-frequency cepstral coefficient feature matrix to obtain corresponding cluster centers corresponding to the valid snoring frames;

[0029] The number of valid snore frames corresponding to each cluster center is determined, and feature data of the corresponding valid snore frames is determined based on the corresponding number and the cluster center.

[0030] Preferably, the real-time apnea and hypopnea prediction method based on snoring, S3: before extracting positive and negative samples using the respiratory event time nodes annotated in the polysomnography data as annotations, comprises:

[0031] Performing quality assessment on the data in the sleep index and the feature data of the valid snoring frames to obtain a complete data sample;

[0032] obtaining a sleep stage of the target subject based on the polysomnography data, and obtaining a time period during which the target subject is not asleep based on the sleep stage;

[0033] Delete the data samples corresponding to the non-sleep time period;

[0034] The retained data samples are fused with the time axis to obtain the data samples covering the entire time domain, and the data samples covering the entire time domain are discretized into equal intervals based on the mean to obtain discrete samples covering the entire time domain.

[0035] Preferably, the real-time apnea and hypopnea prediction method based on snoring, S3: extracting positive and negative samples using the respiratory event time nodes marked in the polysomnography data as labels, includes:

[0036] Acquire time-aligned polysomnographic data;

[0037] Extracting a preset number of positive prediction samples and negative prediction samples based on the time nodes corresponding to the re-annotated respiratory times of the polysomnography data as labels, and extracting the energy distribution ratios of the positive prediction samples, the negative prediction samples, and the valid snoring frames contained in the time nodes as feature vectors of the corresponding valid snoring frames;

[0038] Calculating a first distance difference between each pair of feature vectors corresponding to the valid snoring frames included in the positive prediction sample, calculating a second distance difference between each pair of feature vectors corresponding to the valid snoring frames included in the negative prediction sample, and calculating a third distance difference between the feature vectors corresponding to all valid snoring frames included in the positive prediction sample and the negative prediction sample and the feature vector of the valid snoring frame corresponding to the time node;

[0039] Determine whether the first distance difference and the second distance difference are both less than a preset distance threshold and the third distance difference is greater than the preset distance threshold, and if so, determine that the positive prediction sample and the negative prediction sample meet the requirements;

[0040] Otherwise, when the third distance difference is less than the preset distance threshold, a respiratory event is re-determined; otherwise, when at least one of the first distance difference and the second distance difference is greater than the preset distance difference, valid snoring frames corresponding to the value greater than the preset distance difference are used as positive prediction sample or negative prediction sample selection boundaries;

[0041] The positive prediction samples and the negative prediction samples both include: anthropometric data and sleep information of the target object in a corresponding time period and feature data of valid snoring frames.

[0042] Preferably, the real-time apnea and hypopnea prediction method based on snoring, S4: predicting the probability of apnea and hypopnea occurring within a subsequent preset time based on sleep index data and positive and negative sample extraction results, further includes:

[0043] Extract feature data from a preset number of dimensions for the positive and negative prediction samples based on a sliding window algorithm to obtain a first feature data set corresponding to the positive prediction sample or the negative prediction sample;

[0044] Based on a typical association algorithm, the first feature data of each positive prediction sample or negative prediction sample are fused pairwise to obtain a second feature data set corresponding to the positive prediction sample or negative prediction sample;

[0045] Based on the principle of canonical correlation analysis, the two sets of high-order data are reduced to one dimension, and the correlation coefficients of any two sets of feature data in the feature groups of the positive prediction sample or the negative prediction sample are analyzed;

[0046] Classifying the first feature data based on the correlation coefficient to obtain a classification feature data set corresponding to the positive prediction sample or the negative prediction sample;

[0047] The feature group includes: a first feature data set and a second feature data set.

[0048] Preferably, the real-time apnea and hypopnea prediction method based on snoring, after S obtains the classification feature dataset corresponding to the positive prediction sample or the negative prediction sample, further comprises:

[0049] Inputting the classification feature data set corresponding to the positive prediction sample or the negative prediction sample into the gradient boosting tree model;

[0050] Performing feature importance analysis on the results output by the gradient boosting tree model based on an ensemble learning model to obtain weight values ​​corresponding to features in multiple dimensions;

[0051] Establishing a logistic regression model based on the weight values ​​corresponding to the features of the multiple dimensions and a preset method to obtain scores of the features of the multiple dimensions;

[0052] Comparing the performance of multiple preset models in a preset model library based on the classification feature data set corresponding to the positive prediction sample or the negative prediction sample, and selecting the optimal feature fusion model;

[0053] Inputting the classification feature data set corresponding to the positive prediction sample or the negative prediction sample and the score of the corresponding feature into the optimal feature fusion model to obtain a fusion feature;

[0054] Inputting the fusion features into the integrated model to predict the probability of apnea and hypopnea occurring within a subsequent preset time;

[0055] The probability of apnea and hypopnea occurring within the subsequent preset time is divided into a high probability interval and a low probability interval based on a preset probability threshold.

[0056] Preferably, the real-time apnea and hypopnea prediction method based on snoring, after predicting the probability of apnea and hypopnea occurring within a subsequent preset time, further comprises:

[0057] Set warning thresholds;

[0058] When the probability of apnea and hypopnea occurring within the subsequent preset time is greater than the warning threshold, a warning signal is issued and an intervention action is taken on the target subject;

[0059] Otherwise, the probability of apnea and hypopnea occurring within the subsequent preset time is retained.

[0060] Other features and advantages of the present invention will be described in the following description, and in part will become apparent from the description, or will be understood by practicing the present invention. The purposes and other advantages of the present invention can be realized and obtained by the structures particularly pointed out in the written description, claims, and drawings.

[0061] The technical solution of the present invention is further described in detail below through the accompanying drawings and embodiments. BRIEF DESCRIPTION OF THE DRAWINGS

[0062] 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:

[0063] Figure 1 The figure is a workflow diagram of a real-time apnea and hypopnea prediction method based on snoring in an embodiment of the present invention. DETAILED DESCRIPTION

[0064] The preferred embodiments of the present invention are described below with reference to the accompanying drawings. It should be understood that the preferred embodiments described herein are only used to illustrate and explain the present invention, and are not used to limit the present invention.

[0065] Example 1:

[0066] This embodiment provides a real-time apnea and hypopnea prediction method based on snoring. Figure 1 ,include:

[0067] S1: Obtain the target subject’s snoring audio and sleep index data;

[0068] S2: Preprocessing and feature extraction of the snoring audio;

[0069] S3: Extract positive and negative samples using the respiratory event time nodes annotated in the polysomnography data as annotations;

[0070] S4: Based on sleep index data and positive and negative sample extraction results, predict the probability of apnea and hypopnea within the subsequent preset time.

[0071] In this embodiment, the snoring audio and sleep index data of the target subject are both detected and obtained in real time, wherein the sleep index data includes: polysomnographic data, anthropometric data and sleep information of the target subject, the polysomnographic data is obtained by a polysomnographic monitor, and the snoring audio is obtained and recorded by a recording device.

[0072] The beneficial effects of the above technical solution are: by acquiring the snoring audio and sleep index data of the target object in real time, the prediction data of the respiratory event prediction mechanism is enriched, so that the respiratory event prediction takes into account the multi-faceted influence of the user's own situation and the real-time sleep data of the target object, and improves the prediction accuracy of respiratory events. A respiratory event prediction mechanism based on snoring and other real-time sleep indicators is established, and it is modeled in a unified feature space based on feature fusion and combined with the principle of feature importance, to objectively and accurately predict the occurrence of apnea and hypopnea and their corresponding probabilities. It can monitor before the occurrence node of the respiratory event and issue corresponding warning signals according to the detection results and actively intervene in the target object, realizing data support for the physician's subjective judgment of respiratory changes caused by upper airway changes before and after the respiratory event, and laying a solid algorithm foundation for the subsequent construction of a complete system.

[0073] Example 2:

[0074] Based on the above embodiment 1, this embodiment provides a real-time apnea and hypopnea prediction method based on snoring. S2: preprocessing and feature extraction of the snoring audio, including:

[0075] performing noise reduction processing on the snoring audio based on multi-window spectral subtraction and Wiener filtering, and then performing pre-emphasis processing to obtain a pre-processed snoring audio signal;

[0076] Non-snoring frames and breathing frames are filtered out based on preset hard discrimination and soft discrimination methods.

[0077] In this embodiment, since the processed audio data consists of snoring data from an entire night, which contains a large amount of noise and interference signals, a multi-step preprocessing operation is required. Wiener filtering is applied to the extracted 30-second time sample slices, using a parallel approach of hard and soft discrimination to filter out non-snoring and breathing frames.

[0078] In this embodiment, the multi-window spectral subtraction method can effectively reduce the noise of the snoring audio and perform smoothing between frames.

[0079] In this embodiment, noise reduction is performed on the snoring audio based on the Wiener filter.

[0080] In this embodiment, the pre-emphasis processing is a signal processing method for compensating for the high-frequency components of the snoring audio signal, which is beneficial for enhancing the effective snoring frame signal in the snoring audio signal.

[0081] In this embodiment, the hard discrimination specifically refers to: valid snoring frames with a duration of 0.2-5s, a loudness of 70 decibels or above, or a frequency of 800 Hz or above are treated as abnormal valid snoring frames and are cleaned and removed;

[0082] The soft discrimination specifically refers to: performing cluster analysis on some of the features extracted and selected from a patient, and setting a threshold to eliminate the features accordingly.

[0083] The beneficial effect of the above technical solution is that: by preprocessing the snoring audio, a preprocessed snoring audio signal with a higher signal-to-noise ratio, smoother frame-to-frame transmission, and stronger effective snoring frame signal can be obtained, which is conducive to the subsequent extraction of effective snoring segments in the snoring audio and feature extraction thereof.

[0084] Example 3:

[0085] Based on the above embodiment 2, this embodiment provides a real-time apnea and hypopnea prediction method based on snoring, S2: preprocessing and feature extraction of the snoring audio, further comprising:

[0086] Calculating the background signal threshold of the preprocessed snoring audio signal;

[0087] Dividing the preprocessed snoring audio signal into a plurality of audio segments of a preset length, extracting acoustic features of each audio segment (using the MCFF algorithm to extract from each audio segment), and determining, based on the acoustic features of each audio segment, a candidate region for selecting snoring segment starting points (i.e., a region formed by preliminarily screening the starting points of the snoring segments) and a candidate region for selecting snoring segment ending points (i.e., a region formed by preliminarily screening the ending points of the snoring segments);

[0088] The frame corresponding to the maximum value of the inverse difference function of the starting point candidate area (i.e. the maximum absolute value of the short-time amplitude difference between adjacent audio signal frames) is used as the starting point of the snoring segment (i.e. the starting pair of the snoring segment).

[0089] The frame corresponding to the maximum value of the inverse difference function of the termination point candidate area is used as the termination point of the snoring segment (i.e., the snoring frame corresponding to the termination of the snoring segment);

[0090] determining all snoring segments in the continuous snoring audio signal (audio signal segments preliminarily determined to be snoring) based on adjacent starting points and ending points;

[0091] Obtain the short-time energy and zero-crossing rate (i.e., the number of times the snoring frame signal passes through zero) corresponding to all snoring frames, and determine whether the short-time amplitude of the snoring frame is greater than a minimum energy threshold or the zero-crossing rate is greater than a zero-crossing rate threshold. If so, the snoring frame is determined to be a valid snoring frame; otherwise, the snoring frame is determined to be an invalid snoring frame.

[0092] Obtaining a time interval between adjacent valid snoring frames in the continuous snoring audio signal, and when the time interval is less than a preset time threshold, merging a first valid snoring frame and a second valid snoring frame in the adjacent valid snoring frames to obtain a merged audio signal;

[0093] Otherwise, do not merge.

[0094] In this embodiment, the acoustic features are input into a hybrid neural network model to separate the sound segments, the silent segments, and the mixed segments. The mixed segments are the candidate starting point regions or the candidate snoring segment ending point regions.

[0095] When the audio segment before the mixed audio segment is a silent audio segment and the audio segment after the mixed audio segment is a sound audio segment, determining that the mixed audio segment is a starting point candidate area;

[0096] When the audio segment before the mixed audio segment is a sound audio segment and the audio segment after the mixed audio segment is a silent audio segment, the mixed audio segment is determined to be the end point candidate area.

[0097] In this embodiment, the idea of ​​AV-Box-based endpoint detection is adopted, and a double threshold is set according to local information, so that endpoint detection can achieve better results under specific tasks and time periods.

[0098] In this embodiment, the short-term amplitude is the sum of the absolute values ​​of the audio signal amplitudes of the current frame and all previous frames.

[0099] In this embodiment, the snoring frames are preliminarily screened audio signal frames containing snoring.

[0100] In this embodiment, the maximum value of the inverse difference function represents the maximum absolute value of the short-time amplitude difference between adjacent audio signal frames, indicating that this is the boundary between a voiced frame and a silent frame, or the boundary between a silent frame and a voiced frame.

[0101] The beneficial effect of the above technical solution is: by setting an endpoint detection method with an adaptive threshold and setting a double threshold and hard decision condition, valid snoring segments are accurately screened out, and the valid snoring segments are merged to obtain a more streamlined continuous snoring audio, thereby achieving the simplification of the snoring audio signal while ensuring the validity of the snoring audio signal.

[0102] Example 4:

[0103] Based on the above embodiment 3, this embodiment provides a real-time apnea and hypopnea prediction method based on snoring, S2: preprocessing and feature extraction of the snoring audio, and further comprising:

[0104] Input the combined audio into a deep neural network to extract a Mel-frequency cepstral coefficient feature matrix (a mainstream feature used in speech recognition);

[0105] determining a mutation point of the merged audio (i.e., a point at which the type of the continuous snoring audio changes) based on a first preset criterion (Bayesian Information Criterion), and segmenting the merged audio into a plurality of audio samples based on the mutation point;

[0106] determining an arrangement order of valid snore frames in each audio sample, wherein each audio sample is composed of a preset number of consecutive valid snore frames;

[0107] Clustering valid snoring frames in the same arrangement order in each audio sample based on the Mel-frequency cepstral coefficient feature matrix to obtain corresponding cluster centers corresponding to the valid snoring frames (i.e., clustering according to the cluster centers);

[0108] The number of valid snore frames corresponding to each cluster center is determined, and feature data of the corresponding valid snore frame (ie, including the cluster center and the corresponding number of valid snore frames) is determined based on the corresponding number and the cluster center.

[0109] In this embodiment, extracting the Mel-frequency cepstral coefficient feature includes: performing a discrete Fourier transform on each frame of the audio signal in the continuous snoring audio to obtain a linear spectrum corresponding to each frame of the audio signal;

[0110] Obtaining a Mel frequency spectrum corresponding to each frame of audio signal based on a Mel frequency filter, transforming the Mel frequency spectrum into a cepstral domain, and obtaining a Mel frequency cepstral coefficient feature corresponding to each frame of audio signal;

[0111] The Mel-frequency cepstral coefficient features are arranged in the order of frames to obtain the Mel-frequency cepstral coefficient feature matrix.

[0112] In this embodiment, in the feature extraction part, feature extraction is performed on the respiratory frames in the sample. The time-frequency domain features currently used are mainstream features for speech recognition, such as MFCC and LPCC. After framing in the time domain, inter-frame features are merged using T distribution, stochastic neighbor embedding, template mean, and maximum values. Other features include body position, sleep stage, age, etc., some of which need to be encoded.

[0113] The beneficial effects of the above technical solution are: based on the frame extraction of the Mel-frequency cepstral coefficient feature matrix of the continuous snoring audio, the mutation points are determined based on preset criteria, and cluster analysis is performed and combined to effectively and accurately extract positive and negative samples using the respiratory event time nodes marked in the polysomnography data as annotations.

[0114] Example 5:

[0115] Based on the above-mentioned embodiment 1, this embodiment provides a real-time apnea and hypopnea prediction method based on snoring. S3: Before extracting positive and negative samples using the respiratory event time nodes annotated in the polysomnography data as annotations, the method includes:

[0116] Performing quality assessment on the data in the sleep index and the feature data of the valid snoring frames to obtain a complete data sample;

[0117] obtaining a sleep stage of the target subject based on the polysomnography data, and obtaining a time period during which the target subject is not asleep based on the sleep stage;

[0118] Delete the data samples corresponding to the non-sleep time period;

[0119] The retained data samples and the time axis are fused to obtain the data samples covering the entire time domain (i.e., the retained data samples covering the entire time domain are obtained by combining them with the time axis), and the data samples covering the entire time domain are discretized into equal intervals based on the mean to obtain discrete samples covering the entire time domain.

[0120] In this embodiment, the quality assessment includes: detecting the integrity, validity, and accuracy of the data in the sleep index and the feature data of the valid snoring frames, and evaluating and rejecting the data in the sleep index and the feature data of the valid snoring frames.

[0121] In this embodiment, a fully automated sample extraction algorithm uses the PSG output event list as the gold standard and the respiratory event time nodes annotated in the PSG data as annotations to extract positive and negative samples, achieving full temporal coverage. Physicians only need to spot-check annotated samples and confirm the proportion of snoring frames. This approach significantly reduces the amount of annotation required by clinicians and facilitates the inclusion of new data. By aligning PSG data with audio tracks to simulate real-world conditions, the concepts of stable negative samples and discrete negative samples are proposed, and parameter thresholds are set to ensure a consistent ratio of positive and negative samples.

[0122] In this embodiment: Taking into account the complexity of audio throughout the night, this proposal allows noise reduction through data cleaning based on the original algorithm: for example, removing sample segments when the user is not asleep based on real-time information about sleep stages.

[0123] The beneficial effect of the above technical solution is: based on the acquired sleep stage data, the present invention removes the non-sleep time period in the sample data, achieves accurate cleaning of data samples, and achieves data simplification while ensuring the integrity of the data for subsequent predicted breathing time.

[0124] Example 6:

[0125] Based on the above-mentioned Example 5, this embodiment provides a real-time apnea and hypopnea prediction method based on snoring. S3: extracting positive and negative samples using the respiratory event time nodes annotated in the polysomnography data as annotations, including:

[0126] Acquire time-aligned polysomnographic data;

[0127] Extract a preset number of positive prediction samples (positive samples for the target category corresponding to the true value) and negative prediction samples (negative samples for all other target categories not corresponding to the true value) based on the time nodes corresponding to the re-annotated respiratory times of the polysomnography data as labels, and extract the energy distribution ratio of the valid snoring frames contained in the positive prediction samples, the negative prediction samples, and the time nodes (i.e., the percentage of the spectrum energy of the valid snoring frame to the total spectrum energy of all positive and negative prediction samples) as the feature vector of the corresponding valid snoring frame;

[0128] Calculating a first distance difference between each pair of feature vectors corresponding to the valid snoring frames included in the positive prediction sample, calculating a second distance difference between each pair of feature vectors corresponding to the valid snoring frames included in the negative prediction sample, and calculating a third distance difference between the feature vectors corresponding to all valid snoring frames included in the positive prediction sample and the negative prediction sample and the feature vector of the valid snoring frame corresponding to the time node;

[0129] Determine whether the first distance difference and the second distance difference are both less than a preset distance threshold and the third distance difference is greater than the preset distance threshold, and if so, determine that the positive prediction sample and the negative prediction sample meet the requirements;

[0130] Otherwise, when the third distance difference is less than the preset distance threshold, a respiratory event is re-determined; otherwise, when at least one of the first distance difference and the second distance difference is greater than the preset distance difference, valid snoring frames corresponding to the value greater than the preset distance difference are used as positive prediction sample or negative prediction sample selection boundaries;

[0131] The positive prediction samples and the negative prediction samples both include: anthropometric data and sleep information of the target object in a corresponding time period and feature data of valid snoring frames.

[0132] In this embodiment, valid snoring frames corresponding to a difference greater than the preset distance are used as the positive prediction sample or negative prediction sample selection boundary, that is, all positive samples or negative samples from the valid snoring frames corresponding to a difference greater than the preset distance to the time node are extracted.

[0133] The beneficial effects of the above technical solution are as follows: the present invention designs an automatic sample extraction and annotation algorithm, using the respiratory event time nodes annotated in the PSG data as annotations to extract positive and negative samples, achieving full time domain coverage. Doctors only need to spot-check the annotated samples and confirm the snoring frame ratio, which greatly reduces the amount of annotation required by clinicians and facilitates the inclusion of new data. The present invention also proposes the concepts of stable negative samples and discrete negative samples, and sets parameter thresholds to ensure a relatively consistent ratio of positive and negative samples, achieving accurate extraction of valid samples for predicting respiratory time.

[0134] Example 7:

[0135] Based on the above-mentioned embodiment 1, this embodiment provides a real-time apnea and hypopnea prediction method based on snoring. S4: Based on the sleep index data and the positive and negative sample extraction results, the probability of apnea and hypopnea occurring within a subsequent preset time is predicted, and the following steps are also included:

[0136] Extracting feature data from a preset number of dimensions for the positive and negative prediction samples based on a sliding window algorithm to obtain a first feature data set corresponding to the positive prediction sample or the negative prediction sample (including feature data of the snoring audio and polysomnography data corresponding to the positive prediction sample or the negative prediction sample);

[0137] Based on a typical association algorithm, the first feature data of each positive prediction sample or negative prediction sample are fused in pairs to obtain a second feature data set corresponding to the positive prediction sample or negative prediction sample (a feature data set after two-dimensional feature fusion);

[0138] Based on the principle of canonical correlation analysis, the two sets of high-order data are reduced to one dimension, and the correlation coefficients of any two sets of feature data in the feature groups of the positive prediction sample or the negative prediction sample are analyzed;

[0139] Classifying the first feature data based on the correlation coefficient to obtain a classification feature data set corresponding to the positive prediction sample or the negative prediction sample;

[0140] The preset multiple dimensions include: the target object's body position, sleep stage, heart rate, blood oxygen saturation, respiratory airflow parameters, gender, age, effective snoring frames, etc.

[0141] The feature group includes: a first feature data set and a second feature data set.

[0142] In this embodiment, analyzing the correlation coefficient of any two sets of feature data in the feature groups of the positive prediction sample or the negative prediction sample based on the correlation coefficient includes:

[0143]

[0144] Wherein, ρ(X,Y) is the correlation coefficient of any two sets of feature data in the feature group of the positive prediction sample or the negative prediction sample, X and Y are any two sets of feature data in the feature group of the positive prediction sample or the negative prediction sample, D(X) is the variance of X, D(Y) is the variance of X, and COV(X,Y) is the covariance of X and Y.

[0145] The beneficial effect of the above technical solution is: classifying the positive and negative prediction samples, achieving further simplification and classification of the positive and negative prediction samples, which is conducive to the subsequent determination of the weight value of each type of feature, and thus providing effective data for the subsequent prediction of respiratory events.

[0146] Example 8:

[0147] Based on the above-mentioned embodiment 7, this embodiment provides a real-time apnea and hypopnea prediction method based on snoring. After obtaining the classification feature dataset corresponding to the positive prediction sample or the negative prediction sample, the method further includes:

[0148] Inputting the classification feature data set corresponding to the positive prediction sample or the negative prediction sample into the gradient boosting tree model;

[0149] Performing feature importance analysis on the output of the gradient boosting tree model based on the ensemble learning model to obtain weight values ​​corresponding to features of multiple dimensions (i.e., the output result obtained by inputting the output of the gradient boosting tree model into the ensemble learning model);

[0150] A logistic regression model is established based on the weight values ​​corresponding to the features of the multiple dimensions and a preset method (SelectFromModel method) to obtain the scores of the features of the multiple dimensions (reflecting the influence of the features in each sample and also showing the positive and negative influence);

[0151] Based on the classification feature data set corresponding to the positive prediction sample or the negative prediction sample, a performance comparison is performed on multiple preset models (late fusion, early fusion, slow fusion, etc.) in a preset model library, and an optimal feature fusion model (the feature fusion model most suitable for the present invention) is selected;

[0152] Inputting the classification feature data set corresponding to the positive prediction sample or the negative prediction sample and the corresponding feature score into the optimal feature fusion model to obtain a fusion feature (for obtaining the accuracy of predicting respiratory events);

[0153] Inputting the fusion features into the integrated model to predict the probability of apnea and hypopnea occurring within a subsequent preset time;

[0154] Based on a preset probability threshold, the probability of apnea and hypopnea occurring within the subsequent preset time is divided into a high probability interval and a low probability interval (the output probability of the integrated model is divided into certain intervals: the prediction result of the high probability interval is used as an explanatory result with high model confidence, and the model has a higher confidence in accurate prediction; the prediction result of the low probability interval is used as an explanatory result with low model confidence, and the confidence in accurate model prediction is low).

[0155] In this embodiment: the data used include snoring audio throughout the night, time-aligned PSG data, and the patient's anthropometric data and sleep information. This proposal is committed to integrating these data, flexibly selecting and fusing them. For audio and physical examination data, after using T distribution and stochastic neighbor embedding methods to merge inter-frame information between time domain frames, this proposal uses the SelectFromModel method to establish a logistic regression model for feature scoring; based on the combination of real-time sleep indicators and patient physical examination data, this proposal uses late fusion, early fusion, slow fusion and other models to compare performance and select the feature fusion model that is most suitable for this proposal. In order to improve the interpretability of the model and the demand for model deployment speed, this proposal intends to use XGBoost for data modeling in the python3.8 environment.

[0156] In this embodiment, the focus is on explaining sleep physiological parameters and physical examination characteristics, and further assisting in establishing a subsequent real-time early warning and diagnosis system, which includes:

[0157] Feature Importance Analysis: SHAP was used to analyze the output of the gradient boosting tree model. This model, inspired by cooperative game theory, constructs an additive explanatory model that considers all features as contributors. Its advantage is that SHAP reflects the influence of each feature in a sample, and also indicates whether the influence is positive or negative. By examining the overall SHAP value, the results show that sleep position is a prominent factor among many features.

[0158] Confidence interval warning analysis: The output probability of the integrated model is divided into certain intervals: the prediction results in the high probability interval serve as explanatory results with high model confidence, and the model has a higher degree of confidence in the accurate prediction; the prediction results in the low probability interval serve as explanatory results with low model confidence, and the model has a lower degree of confidence in the accurate prediction.

[0159] The beneficial effects of the above technical solutions are: based on the classification feature data set and the SHAP model, the feature importance of the gradient boosting tree model output is analyzed, reflecting the influence of each feature in the sample, which is beneficial to obtain the output prediction probability corresponding to the integrated model, and divide the subsequent preset interval into a high probability interval and a low probability interval, which can accurately determine whether to remind, effectively improve the situation of triggering respiratory events.

[0160] Embodiment 9:

[0161] Based on the above embodiment 8, the embodiment provides a snoring sound-based real-time apnea and hypopnea prediction method, after predicting the probability of occurrence of apnea and hypopnea in the subsequent preset time, it further includes:

[0162] Setting an early warning threshold (pre-set according to actual conditions);

[0163] When the probability of occurrence of apnea and hypopnea in the subsequent preset time is greater than the early warning threshold, an early warning signal is sent and an intervention action is taken on the target object (such as using hypoglossal nerve stimulation, body position intervention method, etc.);

[0164] Otherwise, the probability of occurrence of apnea and hypopnea in the subsequent preset time is retained.

[0165] The beneficial effects of the above technical solutions are: by pre-setting the early warning threshold according to the actual situation, it can automatically determine whether to intervene and remind, and realize avoiding too much disturbance to the user's sleep while effectively improving the situation of triggering respiratory events.

[0166] Obviously, those skilled in the art can make various modifications and variations to the present application without departing from the spirit and scope of the present application. Thus, if these modifications and variations of the present application fall within the scope of the claims of the present application and their equivalent technologies, the present application also intends to include these modifications and variations.

Claims

1. A real-time apnea and hypopnea prediction method based on snoring, characterized in that: include: S1: Obtain the target subject’s snoring audio and sleep index data; S2: Preprocessing and feature extraction of the snoring audio; S3: Extract positive and negative samples using the respiratory event time nodes annotated in the polysomnography data as annotations; S4: Based on sleep index data and positive and negative sample extraction results, predict the probability of apnea and hypopnea within the subsequent preset time; Among them, S4: Based on sleep indicator data and positive and negative sample extraction results, predict the probability of apnea and hypopnea within the subsequent preset time, and also includes: Extract feature data from a preset number of dimensions for the positive and negative prediction samples based on a sliding window algorithm to obtain a first feature data set corresponding to the positive prediction sample or the negative prediction sample; Based on a typical association algorithm, the first feature data of each positive prediction sample or negative prediction sample are fused in pairs to obtain a second feature data set corresponding to the positive prediction sample or negative prediction sample; Based on the principle of canonical correlation analysis, the two sets of high-dimensional data are reduced to one dimension, and the correlation coefficients of any two sets of feature data in the feature groups of the positive prediction sample or the negative prediction sample are analyzed; Classifying the first feature data based on the correlation coefficient to obtain a classification feature data set corresponding to the positive prediction sample or the negative prediction sample; Wherein, the feature group includes: a first feature data set and a second feature data set; After obtaining the classification feature dataset corresponding to the positive prediction sample or the negative prediction sample, the method further includes: Inputting the classification feature data set corresponding to the positive prediction sample or the negative prediction sample into the gradient boosting tree model; Performing feature importance analysis on the results output by the gradient boosting tree model based on an ensemble learning model to obtain weight values ​​corresponding to features in multiple dimensions; Establishing a logistic regression model based on the weight values ​​corresponding to the features of the multiple dimensions and a preset method to obtain scores of the features of the multiple dimensions; Comparing the performance of multiple preset models in a preset model library based on the classification feature data set corresponding to the positive prediction sample or the negative prediction sample, and selecting the optimal feature fusion model; Inputting the classification feature data set corresponding to the positive prediction sample or the negative prediction sample and the score of the corresponding feature into the optimal feature fusion model to obtain a fusion feature; Inputting the fusion features into the integrated learning model to predict the probability of apnea and hypopnea occurring within a subsequent preset time; Dividing the probability of apnea and hypopnea occurring within the subsequent preset time into a high probability interval and a low probability interval based on a preset probability threshold; After predicting the probability of apnea and hypopnea within the subsequent preset time, it also includes: Set warning thresholds; When the probability of apnea and hypopnea occurring within the subsequent preset time is greater than the warning threshold, a warning signal is issued and an intervention action is taken on the target subject; Otherwise, the probability of apnea and hypopnea occurring within the subsequent preset time is retained.

2. The method for real-time apnea and hypopnea prediction based on snoring according to claim 1, characterized in that: S2: Preprocessing and feature extraction of the snoring audio, including: performing noise reduction processing on the snoring audio based on multi-window spectral subtraction and Wiener filtering, and then performing pre-emphasis processing to obtain a pre-processed snoring audio signal; Based on preset hard discrimination and soft discrimination methods, non-snoring frames and breathing frames in the pre-processed snoring audio signal are filtered out.

3. The method for real-time apnea and hypopnea prediction based on snoring according to claim 2, characterized in that: S2: Preprocessing and feature extraction of the snoring audio, further comprising: Calculating the background signal threshold of the preprocessed snoring audio signal; Dividing the preprocessed snoring audio signal into a plurality of audio segments of a preset length, extracting acoustic features of each audio segment, and determining a candidate region for selecting a snoring segment start point and a candidate region for selecting a snoring segment end point based on the acoustic features of each audio segment; The frame corresponding to the maximum value of the inverse difference function of the starting point candidate area is used as the starting point of the snoring segment, and the frame corresponding to the maximum value of the inverse difference function of the ending point candidate area is used as the ending point of the snoring segment; determining all snoring segments in the continuous snoring audio signal based on adjacent starting points and ending points; Obtaining the short-time energy and zero-crossing rate corresponding to all snoring frames, and determining whether the short-time amplitude of the snoring frame is greater than a minimum energy threshold or the zero-crossing rate is greater than a zero-crossing rate threshold; if so, determining the snoring frame as a valid snoring frame; otherwise, determining the snoring frame as an invalid snoring frame; Obtaining a time interval between adjacent valid snoring frames in the continuous snoring audio signal, and when the time interval is less than a preset time threshold, merging a first valid snoring frame and a second valid snoring frame in the adjacent valid snoring frames to obtain a merged audio signal; Otherwise, do not merge.

4. The method for real-time apnea and hypopnea prediction based on snoring according to claim 3, wherein: S2: Preprocessing and feature extraction of the snoring audio, further comprising: Inputting the combined audio into a deep neural network to extract a Mel-frequency cepstral coefficient feature matrix; Determining a mutation point of the merged audio based on a first preset criterion, and dividing the merged audio into a plurality of audio samples based on the mutation point; determining an arrangement order of valid snore frames in each audio sample, wherein each audio sample is composed of a preset number of consecutive valid snore frames; Clustering valid snoring frames in the same arrangement order in each audio sample based on the Mel-frequency cepstral coefficient feature matrix to obtain cluster centers corresponding to the valid snoring frames; The number of valid snore frames corresponding to each cluster center is determined, and feature data of the corresponding valid snore frames is determined based on the corresponding number and the cluster center.

5. The method for real-time apnea and hypopnea prediction based on snoring according to claim 1, characterized in that: S3: Before extracting positive and negative samples using the respiratory event time nodes annotated in the polysomnography data as annotations, including: Performing quality assessment on the data in the sleep index and the feature data of the valid snoring frames to obtain a complete data sample; obtaining a sleep stage of the target subject based on the polysomnography data, and obtaining a time period during which the target subject is not asleep based on the sleep stage; Delete the data samples corresponding to the non-sleep time period; The retained data samples are fused with the time axis to obtain data samples covering the entire time domain, and the data samples covering the entire time domain are discretized into equal intervals based on the mean to obtain discrete samples covering the entire time domain.

6. The method for real-time apnea and hypopnea prediction based on snoring according to claim 5, characterized in that: S3: Extract positive and negative samples using the respiratory event time nodes annotated in the polysomnography data as annotations, including: Acquire time-aligned polysomnographic data; Based on the time nodes corresponding to the re-annotated breathing times of the polysomnography data as labels, a preset number of positive prediction samples and negative prediction samples are extracted, and the energy distribution ratios of the valid snoring frames contained in the positive prediction samples, the negative prediction samples, and the time nodes are extracted as feature vectors of the corresponding valid snoring frames; Calculating a first distance difference between each pair of feature vectors corresponding to the valid snoring frames included in the positive prediction sample, calculating a second distance difference between each pair of feature vectors corresponding to the valid snoring frames included in the negative prediction sample, and calculating a third distance difference between the feature vectors corresponding to all valid snoring frames included in the positive prediction sample and the negative prediction sample and the feature vector of the valid snoring frame corresponding to the time node; Determine whether the first distance difference and the second distance difference are both less than a preset distance threshold and the third distance difference is greater than the preset distance threshold, and if so, determine that the positive prediction sample and the negative prediction sample meet the requirements; Otherwise, when the third distance difference is less than the preset distance threshold, it is determined that a respiratory event is re-determined; otherwise, when at least one of the first distance difference and the second distance difference is greater than the preset distance threshold, valid snoring frames corresponding to values ​​greater than the preset distance threshold are used as positive prediction sample or negative prediction sample selection boundaries; The positive prediction samples and the negative prediction samples both include: anthropometric data and sleep information of the target object in a corresponding time period and feature data of valid snoring frames.

Images