A sleep feature extraction method based on probability density function and polysomnogram

Abstract

The application discloses a sleep feature extraction method based on a probability density function and a polysomnogram. The method comprises collecting and preprocessing PSG electrophysiological signal data of a subject, marking sleep stages of the PSG data; dividing the preprocessed PSG data into time window frames with a fixed time length, extracting time window features in the time domain, the frequency domain and the time-frequency domain in each channel, and grouping the features according to different sleep stages; then, calculating the joint probability density function of the time window features in each sleep stage by a kernel density estimation method, and extracting comprehensive statistical features including time features and space features based on the joint probability density function; finally, predicting the age of the subject through the equal comprehensive statistical features. The method can extract stable sleep features from non-stable electrophysiological signals, mine sleep features and visually display the sleep features, so as to be applied to the analysis of various sleep physiological phenomena and sleep disorders.

Description

TECHNICAL FIELD

[0001] The application belongs to the field of sleep data analysis, and particularly relates to a sleep feature extraction method based on a probability density function and a polysomnogram. BACKGROUND

[0002] Sleep disorders affect most of the population, and it is not only a personal problem, but also a social problem. At present, the research means related to sleep disorder diagnosis and its correlation with other major brain diseases are limited, mainly relying on subjective scales, lacking objective and quantitative analysis methods, theories and systems. Polysomnogram (PSG) is a clinical means for objectively and quantitatively recording sleep, which can provide reliable basis for the diagnosis of various sleep diseases, and it contains multi-channel electrophysiological signals. However, it is difficult to extract stable disease-related markers from these non-stable signals, and one of the reasons is that the sleep features of individuals have great heterogeneity. This individual difference is an important problem in sleep science, and describing this individual difference is the premise of mining sleep disorder markers.

[0003] Many studies have pointed out that sleep spindles can characterize the stability of individual sleep, and their occurrence rate has great inter-individual difference. However, the quality of such analysis results is subject to the extraction accuracy of sleep spindles. Stokes et al. expanded the analysis range of spindles to general transient oscillation events, obtained a highly heterogeneous representation between subjects, and distinguished schizophrenic patients from healthy controls based on this representation. However, they did not establish a general feature extraction method for sleep disorders and their comorbidities, and did not convert this inter-individual heterogeneity representation into a mineable feature vector. Gennaro et al. found that the macroscopic spatial distribution of non-REM EEG signals can characterize the individual sleep features, but did not conduct quantitative analysis and association with sleep disorders and their comorbidities. The above studies on stable sleep features provide important clues for understanding individual sleep stability characteristics, but lack a generally applicable feature extraction method, and have limitations in practical application scenarios. On the other hand, there is still a lack of a general sleep stability feature visualization method. Therefore, there is an urgent need for a general method that can overcome the non-stable characteristics of PSG electrophysiological signals, extract individual stable sleep features from them, and convert them into mineable feature vectors and visualization, providing reliable quantitative basis for sleep-related disease auxiliary diagnosis, prognosis, sleep mechanism, and sleep mechanism research. SUMMARY

[0004] In order to solve the problems in the background art, the purpose of the present application is to provide a sleep feature extraction method based on a probability density function and a polysomnogram.

[0005] The technical scheme adopted by the present application comprises the following steps:

[0006] Step S1), collecting polysomnography (PSG) data of a subject and preprocessing, and performing sleep stage annotation on the preprocessed PSG data;

[0007] Step S2), dividing the preprocessed PSG data into fixed-length time windows, and extracting a set of time window features in each polysomnography (PSG) channel for each time window, each set of time window features including time domain features, frequency domain features, and time-frequency domain features;

[0008] Step S3), grouping the time window features in the set of time window features by sleep stage, and estimating the joint probability density function of the time window features on all PSG channels in each sleep stage;

[0009] Step S4), visualizing the joint probability density function obtained in step S3) using a contour map;

[0010] Step S5), obtaining comprehensive statistical features of the polysomnography (PSG) based on the joint probability density function, the comprehensive statistical features including basic statistical features, intra-channel temporal features, and inter-channel spatial features, and further predicting the age of the subject through the comprehensive statistical features of the polysomnography (PSG).

[0011] The polysomnography (PSG) data collected in step S1) includes multi-channel electrophysiological signals, including at least two electroencephalogram (EEG) channels, and other channels including electrooculogram (EOG), electrocardiogram (ECG), and electromyogram (EMG), etc. To ensure that these electrophysiological signals can reflect physiological responses on a microscopic time scale, the minimum sampling rate of each channel is 100 Hz, and the polysomnography (PSG) data is preprocessed to reduce noise in the multi-channel electrophysiological signals. The multi-channel electrophysiological signals include multi-channel electroencephalogram (EEG) signals, multi-channel electrocardiogram (ECG) signals, and multi-channel electromyogram (EMG) signals, etc.

[0012] The polysomnography (PSG) data of the subject is collected and preprocessed, and the preprocessing is specifically: the electroencephalogram (EEG) and electrooculogram (EOG) signals (if present) are subjected to low-frequency filtering at a frequency of 0.3 Hz and high-frequency filtering at a frequency of 35 Hz; the electromyogram (EMG) signal (if present) is subjected to low-frequency filtering at a frequency of 10 Hz and high-frequency filtering at a frequency of 100 Hz; and the electrocardiogram (ECG) signal (if present) is subjected to low-frequency filtering at a frequency of 0.3 Hz and high-frequency filtering at a frequency of 70 Hz, the purpose of preprocessing being to reduce noise in the signals.

[0013] The specific manner of sleep stage annotation of the preprocessed PSG data in the step S1) is: according to the AASM standard of the American Sleep Medicine Association, the preprocessed PSG data is divided into standard sleep frames of 30s, and each standard sleep frame is classified into five different sleep stages of Wake, N1, N2, N3 and REM.

[0014] The advantage of sleep stage annotation is that the statistical characteristics of PSG electrophysiological signals in each sleep stage are relatively stable, which can effectively reduce the complexity of directly analyzing unstable electrophysiological signals in the whole night PSG data.

[0015] The duration of the standard sleep frame is a multiple of the duration of the time window divided in the step S2), so that each standard sleep frame contains one or more time windows. The advantage is that each time window (short time frame) thus divided can be clearly classified into the corresponding sleep stage according to the standard sleep frame to which it belongs.

[0016] In the step S2), different time window feature groups are extracted for different categories of PSG channels. The functional expression of the time window feature group is as follows:

[0017]

[0018] Wherein, represents the mth channel c m corresponding time window feature group; represents the kth time window feature under the channel c m , maps the channel c m in the time window to a scalar, i.e. a time window feature; c m is the identification of the PSG channel; m is the serial number of the channel; is the total number of time window features corresponding to the channel c m , and k is the serial number of the time window feature corresponding to the channel c m .

[0019] The advantage is that each category of channel can extract corresponding time window features according to its electrophysiological signal characteristics. For example, the time window feature group corresponding to the EEG channel tends to include time domain features, frequency domain features and time-frequency domain features; the time window feature group corresponding to the EMG channel tends to include time domain features; and the time window feature group corresponding to the ECG channel tends to include frequency domain features or time-frequency domain features.

[0020] In step (3), the joint probability density estimation is only performed on the time window features of non-Wake stage, i.e., the joint probability density estimation is performed on the time window features of non-rapid eye movement stage 1 (N1), non-rapid eye movement stage 2 (N2), non-rapid eye movement stage 3 (N3) and rapid eye movement stage (REM). The joint probability density function p(X) of the overall time window features of PSG can be obtained from the joint probability density functions p s (X) of the time window features of each stage according to the total probability formula as follows:

[0021] p(X) = ∑ s∈Sd p s (X)·P(s)

[0022]

[0023] Sd = {N1, N2, N3, REM}

[0024] wherein p(X) represents the joint probability density function of the time window features of all PSG channels; s represents the sleep stage, and Sd represents the set of the four sleep stages except for the Wake stage; p s (X) represents the joint probability density function of the time window features under the sleep stage s; P(s) represents the probability that the corresponding time window features belong to the sleep stage s; represents the kth time window feature under the channel c m ; M represents the total number of channels; and X represents the set of time window features of all channels.

[0025] Specifically, the joint probability density function p(X) thus estimated contains (i) the statistical parameters (including the mean and variance) of each time window feature (on its corresponding channel) in each sleep stage; (ii) the statistical parameter offset of each time window feature (on its corresponding channel) in different sleep stages, i.e., the time feature of the time window feature in the corresponding sleep stage, such as the difference in the average amplitude of the signal of a certain EEG channel between N2 stage and N3 stage; (iii) the statistical parameter offset of the same type of electrophysiological signal on different channels, i.e., the spatial feature of the time window feature between the corresponding channels, such as the difference in the average amplitude of the EEG signal channel F4-M1 and the channel C4-M1 in N2 stage; and (iv) the correlation coefficient of any channel and different time window features in each sleep stage.

[0026] For the sleep stage s, the joint probability density function p s (X) of the time window features is estimated by the kernel density estimation method, and the probability P(s) that the time window belongs to the sleep stage s is obtained by processing the following formula:

[0027]

[0028] where |s| denotes the standard sleep frame number of sleep stage s, and denominator ∑ s′∈Sd |s| denotes the total sleep frame number.

[0029] Within each sleep stage, the statistical features of the electrophysiological signals are relatively stable, so the corresponding time window features can be approximately regarded as independent and identically distributed data, and thus are suitable for estimation by kernel density estimation method.

[0030] The step S4) specifically includes:

[0031] Step S4.1), first, a pair of time window features including two different time window features is selected For sleep stage s, the joint probability density function p s (X) of the time window features is processed according to the following formula:

[0032]

[0033] where X' represents all time window features except ; and represents the i-th time window feature under channel c p ; and represents the j-th time window feature under channel c q .

[0034] Step S4.2), an isogram of the edge distribution probability density function is drawn with the time window feature as the horizontal coordinate and the time window feature as the vertical coordinate, as the edge distribution probability density function isogram of the time window feature pair under sleep stage s.

[0035] Step S4.3), steps S4.1) to S4.2) are repeated until the edge distribution probability density function isograms of all sleep stages in the sleep stage set Sd corresponding to the time window feature pair are drawn in the same coordinate system, and then the time window feature pair is visually displayed through the edge distribution probability density function isograms, as shown in Figure 2 .

[0036] The acquisition method of the comprehensive statistical features of the polysomnogram PSG in the step S5) is as follows:

[0037] Step S5.1), first, basic statistical features are extracted, including:

[0038] ​the mean of the time window features in each sleep stage, in each channel;

[0039] the standard deviation of the time window features in each sleep stage, in each channel;

[0040] the correlation coefficient between the time window features in each sleep stage, in each channel;

[0041] The step S5.2) selects the non-rapid eye movement stage 2 N2 as the reference sleep stage, and the N2 stage as the reference sleep stage. The advantage of selecting the N2 stage as the reference sleep stage is that, in the whole night sleep, the N2 stage usually corresponds to the most time, and the statistical characteristics are the most stable. The time features in the channel are extracted, including:

[0042] the difference between the mean of the time window features in each sleep stage, in each channel, and the mean of the time window features in the non-rapid eye movement stage 2 N2;

[0043] the difference between the standard deviation of the time window features in each sleep stage, in each channel, and the standard deviation of the time window features in the non-rapid eye movement stage 2 N2;

[0044] the difference between the correlation coefficient between the time window features in each sleep stage, in each channel, and the correlation coefficient between the time window features in the non-rapid eye movement stage 2 N2;

[0045] The step S5.3) selects a reference channel from all channels for each PSG channel class containing multiple channels, and extracts the inter-channel spatial features of this class of channels based on the selected reference channel, including:

[0046] the difference between the mean of the time window features in each sleep stage, in the remaining channels except the reference channel, and the mean of the time window features in the reference channel;

[0047] the difference between the standard deviation of the time window features in each sleep stage, in the remaining channels except the reference channel, and the standard deviation of the time window features in the reference channel;

[0048] the difference between the correlation coefficient between the time window features in each sleep stage, in the remaining channels except the reference channel, and the correlation coefficient between the time window features in the reference channel.

[0049] The step S5) predicts the age of the subject by comprehensively analyzing the statistical characteristics in the following manner:

[0050] First, an age prediction data set is constructed, including comprehensive statistical features of polysomnography (PSG) and corresponding real ages, and the real ages are taken as labels of the age prediction data set; then, an age prediction model is constructed based on a machine learning feature dimension reduction algorithm and a machine learning regression algorithm, the age prediction data set is input into the constructed age prediction model for training, and a trained age prediction model is obtained; finally, the polysomnography (PSG) data of a subject to be tested are input into the trained age prediction model, and an age prediction result of the subject to be tested is output.

[0051] The comprehensive statistical features of the PSG data include basic statistical features, time features in a channel, and spatial features between channels.

[0052] The above step calculates the relative deviation of the statistical parameters of the feature distribution of each time window based on the estimated kernel density function, and the advantage is that the influence of the potential unstable baseline of the individual electrophysiological characteristics recorded by the PSG on feature extraction is eliminated.

[0053] The present application obtains features with relatively stable statistical distribution by conditionally decomposing the current time window of the PSG electrophysiological signal based on different sleep stages. These features are further extracted as feature vectors representing individual stable sleep features, and the proposed method can be extended to various sleep physiological and sleep disorder analysis applications, including sleep age research and auxiliary diagnosis of sleep disorders, to effectively solve the problems mentioned in the background art.

[0054] The present application has the following advantages:

[0055] 1. The present application divides the whole night PSG data according to sleep stages, estimates the probability density function of the time window features in each stage, overcomes the difficulty of extracting stable features from unstable electrophysiological signals, and can be applied to sleep age research.

[0056] 2. The method of the present application can extract feature vectors from general PSG data for data mining by downstream machine learning algorithms, and has strong universality. Visualizing sleep features can provide intuitive basis for sleep disease research and clinical diagnosis, and provide reliable quantitative basis for auxiliary diagnosis, prognosis, sleep mechanism and sleep mechanism research of sleep-related diseases. BRIEF DESCRIPTION OF DRAWINGS

[0057] Figure 1 The flowchart of the present application;

[0058] Figure 2 The schematic diagram of visualizing sleep features of the present application. DETAILED DESCRIPTION

[0059] The application will be described in further detail below with reference to the drawings and specific embodiments. Obviously, the described embodiments are only a part of the embodiments of the application, rather than all the embodiments. Based on the embodiments in the application, all other embodiments obtained by a person of ordinary skill in the art without creative work fall within the protection scope of the application.

[0060] The embodiment collects a sleep cassette (SC) data subset in the international publication PSG data set Sleep-EDFx (hereinafter referred to as the SC data set).

[0061] The embodiment uses the stable sleep feature extraction method based on the probability density function of the application to perform feature extraction on all PSG data in the SC data set, and then performs individual age prediction based on the extracted features. On the other hand, the embodiment visualizes part of the data in the SC data set.

[0062] As shown in Figure 1 The feature extraction method of the application specifically includes the following steps:

[0063] Step (1): Collecting the polysomnography PSG data of the subject and performing preprocessing, and performing sleep stage labeling on the preprocessed PSG data.

[0064] The embodiment uses the sleep stage data labeled by professional sleep technicians according to the R&K rules included in the SC data set. In order to make the sleep stage labeling comply with the AASM standard, the embodiment classifies the N4 stage as the N3 stage. In this way, each standard sleep frame is classified into five different sleep stages of Wake, N1, N2, N3, and REM.

[0065] The embodiment uses the EEG channels in the SC data set, i.e., the Fpz-Cz and Pz-Oz channels (denoted as c1 and c2, respectively), and the sampling rate is 100 Hz.

[0066] The polysomnography PSG data collected in step S1) contains multi-channel electrophysiological signals, including at least two electroencephalogram EEG channels, and other channels including electrooculogram (EOG), electrocardiogram (ECG), and electromyogram (EMG), etc. In order to ensure that these electrophysiological signals can reflect the physiological response on the micro time scale, the minimum sampling rate of each channel is 100 Hz, and the preprocessing of the polysomnography PSG data is used to reduce the noise in the multi-channel electrophysiological signals. The multi-channel electrophysiological signals include multi-channel electroencephalogram signals, multi-channel electrocardiogram signals, and multi-channel electromyogram signals, etc.

[0067] The PSG data of the subject is collected and preprocessed, and the preprocessing specifically includes: performing low-frequency filtering at a frequency of 0.3 Hz and high-frequency filtering at a frequency of 35 Hz on electroencephalogram (EEG) and electrooculogram (EOG) signals (if present); performing low-frequency filtering at a frequency of 10 Hz and high-frequency filtering at a frequency of 100 Hz on electromyogram (EMG) signals (if present); and performing low-frequency filtering at a frequency of 0.3 Hz and high-frequency filtering at a frequency of 70 Hz on electrocardiogram (ECG) signals (if present), the purpose of the preprocessing being to reduce noise in the signals.

[0068] The specific manner of sleep stage annotation of the preprocessed PSG data in step S1) is: according to the American Academy of Sleep Medicine (AASM) standard, the preprocessed PSG data is divided into standard sleep frames of 30 seconds, and each standard sleep frame is classified into five different sleep stages of wakefulness (Wake), non-rapid eye movement stage 1 (N1), non-rapid eye movement stage 2 (N2), non-rapid eye movement stage 3 (N3), and rapid eye movement stage (REM).

[0069] The advantage of sleep stage annotation is that the statistical characteristics of PSG electrophysiological signals in each sleep stage are relatively stable, which can effectively reduce the complexity of directly analyzing unstable electrophysiological signals in the whole-night PSG data.

[0070] Step (2): The preprocessed PSG data is divided into time windows of a fixed length, and in each PSG channel, a set of time window features is extracted for each time window, each set of time window features including time domain features, frequency domain features, and time-frequency domain features.

[0071] In a specific implementation, the PSG data is divided into time windows (i.e., short time frames) of a fixed length, and in each PSG channel, a set of time window features is extracted for each time window, including time domain features, frequency domain features, and time-frequency domain features.

[0072] The length of the standard sleep frame is a multiple of the length of the time window divided in step S2), so that each standard sleep frame contains one or more time windows. Each time window (short time frame) thus divided can be clearly classified into the corresponding sleep stage according to the standard sleep frame to which it belongs. In this embodiment, each PSG data in the SC data set is divided into time windows of 30 seconds, so that each time window corresponds to a standard sleep frame, i.e., a sleep stage.

[0073] The time window feature group extracted from the brain electrical channel in this embodiment includes 7 time window features, which are: ① time window average frequency; ② time window average amplitude; ③ time window relative power in the Delta frequency band (0.5-4 Hz); ④ time window relative power in the Theta frequency band (4-8 Hz); ⑤ time window relative power in the Alpha frequency band (8-12 Hz); ⑥ time window relative power in the Beta frequency band (12-30 Hz); and ⑦ time window total power in the full frequency band (0.5-30 Hz).

[0074] In this way, in this embodiment, 14-dimensional time window features are extracted for each time window, denoted as

[0075] Step (3), group the time window features in the time window feature group according to the sleep stage, and estimate the joint probability density function of the time window features on all PSG channels in each sleep stage;

[0076] As shown in Figure 1 , this embodiment divides the time window frames of each PSG data in the SC data set into five categories, which correspond to the five different sleep stages of Wake, N1, N2, N3, and REM. For each sleep stage s∈Sd excluding the Wake period, the joint probability density function p s (X) of X in sleep stage s is estimated by the kernel density estimation method. The probability corresponding to the sleep stage s is estimated by the following formula:

[0077]

[0078] where |s| represents the standard sleep frame number of sleep stage s. In this way, the PSG overall time window feature joint probability density function p(X) can be represented as:

[0079] p(X)=∑ s∈Sd p s (X)·P(s)

[0080] Step (4): obtain the comprehensive statistical features of the polysomnogram PSG based on the joint probability density function, and the comprehensive statistical features include basic statistical features, intra-channel temporal features, and inter-channel spatial features;

[0081] In this embodiment, the extracted basic statistical features include, in each sleep stage and each channel, ① the mean of each time window feature, a total of 56 features; ② the standard deviation of each time window feature, a total of 56 features; and ③ the correlation coefficient between each pair of time window features, a total of 168 features. In this embodiment, a total of 280 basic statistical features are extracted.

[0082] In this embodiment, N2 stage is selected as the reference sleep stage, and for each time window feature in the time window feature set X, the intra-channel time features are extracted, including ① the mean offset of each time window feature in each channel relative to the N2 stage, a total of 42 features; ② the standard deviation offset of each time window feature in each channel relative to the N2 stage, a total of 42 features; ③ the difference between the correlation coefficients of each time window feature in each channel relative to the N2 stage, a total of 126 features. In this embodiment, a total of 210 intra-channel time features are extracted.

[0083] In this embodiment, Fpz-Cz is selected as the reference channel, and the inter-channel spatial features are extracted, including ① the mean offset of each time window feature in the Pz-Oz channel relative to the Fpz-Cz channel, a total of 28 features; ② the standard deviation offset of each time window feature in the Pz-Oz channel relative to the Fpz-Cz channel, a total of 28 features; ③ the correlation offset between each time window feature in the Pz-Oz channel relative to the Fpz-Cz channel, a total of 84 features. In this embodiment, a total of 140 inter-channel spatial features are extracted.

[0084] Step (5), based on the comprehensive statistical features, an age prediction data set is constructed, the age prediction model is trained using the age prediction data set, a trained age prediction model is obtained, and finally the trained age prediction model is used to obtain the age prediction result of the subject:

[0085] In this embodiment, for each PSG data in the SC data set, a total of 630-dimensional comprehensive statistical features are extracted from the joint probability density function p(X) of the time window features, and a feature vector set corresponding to the comprehensive statistical features is formed as an age prediction data set.

[0086] Based on the feature vector set and its corresponding true age (hereinafter referred to as CA), this embodiment combines machine learning feature dimension reduction algorithm and machine learning regression algorithm to construct an age prediction model, the input of the model is the feature vector, and the output of the model is the individual age value (hereinafter referred to as PA) predicted by the feature vector. By calculating the mean absolute error (mean absolute error, hereinafter referred to as MAE) and the Pearson correlation coefficient between the PA and the CA corresponding to the SC data set, the performance of the age prediction model construction method can be evaluated.

[0087] Specifically, the machine learning feature dimension reduction algorithm used in this embodiment is the p-value screening method, and the machine learning regression algorithm used is ElasticNet. This embodiment evaluates the performance of the age prediction model construction method in the traditional nested 5-fold cross-validation (nested 5-fold cross-validation) framework. The nested 5-fold cross-validation is randomly and uniformly divided according to CA during data division. For each fold data, the age prediction model is constructed on the training set corresponding to the fold data, and the age prediction is performed on the test set corresponding to the fold data, and the specific steps are as follows:

[0088] Step (1): The feature vector in the training set is reduced to k dimensions by the p-value screening method, and the specific steps are as follows:

[0089] Step (1.1): All features are sorted according to their linear correlation significance (p-value) with age from small to large;

[0090] Step (1.2): The feature index i is set to 2;

[0091] Step (1.3): If the correlation between the i-th feature and one of the first i-1 features is greater than 0.9, the feature is removed; otherwise, the feature index is incremented, i.e. i = i + 1;

[0092] Step (1.4): Repeat step (1.3) until the correlation between each pair of features is less than 0.9;

[0093] Step (1.5): Select the first k features from the remaining features for subsequent calculation;

[0094] Step (2): Based on the training set after feature dimension reduction, search for the optimal configuration of the machine learning ElasticNet through internal 5-fold cross-validation method. Fit the ElasticNet with the optimal configuration on the training set after feature dimension reduction;

[0095] Step (3): In the test set, reduce the dimension of the test set features by selecting the k features obtained in step (1.5); use the ElasticNet fitted in step (2) to predict the test set, and obtain the PA of the test set corresponding to the fold data.

[0096] Finally, the PA of the test set corresponding to each fold data is combined into the PA of the SC data set, and the MAE and Pearson correlation coefficient are calculated based on the PA.

[0097] In this embodiment, the performance of the age prediction models constructed based on three sleep features is compared. The three sleep features are: ① 630-dimensional comprehensive statistical features based on probability density function proposed by the present application; ② 30-dimensional macro sleep features, including the proportion of Wake, N1, N2, N3 and REM period in the time from the first sleep period to the last sleep period (a total of 5 features), and the transition probability from Wake, N1, N2, N3 and REM period to Wake, N1, N2, N3 and REM period (a total of 25 features); and ③ 660-dimensional fusion features combining ① and ②.

[0098] In this embodiment, the above-mentioned nested 5-fold cross-validation algorithm is used to construct the age prediction model for the three sleep features, and only the k value selection of the p value screening method is different. The k value selection for dimension reduction of the 30-dimensional macro sleep features is 10, 20 and 30; and the k value selection for dimension reduction of the 630-dimensional comprehensive statistical features and the 660-dimensional fusion features is 50, 100, 200 and 300.

[0099] In this embodiment, for each feature, 30 repeated experiments are performed for each selectable k value. Table 1 shows the optimal average MAE for each sleep feature in the age prediction task, as well as the corresponding feature dimension k after dimension reduction and the average Pearson correlation coefficient. It can be seen that the comprehensive statistical features based on probability density function extracted by the present application contain more features that can reflect the stable characteristics of individuals than the macro sleep features. On the other hand, the age prediction model constructed based on the fusion features of the two can achieve better age prediction performance.

[0100] Table 1

[0101] Macro-sleep features Integrated statistical features Fusion features Dimension k of features after dimension reduction 20 300 300 MAE 13.35±0.10 8.36±0.21 7.92±0.14 Pearson correlation coefficient 0.68±0.01 0.87±0.01 0.89±0.00

[0102] The second aspect of the present application provides a sleep feature visualization method, which is specifically as follows:

[0103] Based on the joint probability density function p s (X), s∈Sd, select the ① time window average frequency and ② time window average amplitude of the Fpz-Cz channel, i.e. and For each sleep stage, the p s (X) is calculated as the marginal distribution probability density function of

[0104] In this embodiment, 4 healthy people are selected from the SC data set, each with 2 nights of PSG data, for a total of 8 PSG data. For each PSG data, the contour plot of p is displayed on the same coordinate axis, as shown in Figure 2 ​The probability density function of the Wake period is also shown in FIG. 6 Figure 2 The probability density function of the Wake period is also shown in FIG. 6

[0105] The embodiments of the present application introduce a general method of extracting individual stable sleep features from non-stable PSG electrophysiological signals, converting the individual stable sleep features into mineable feature vectors, and visualizing the same. The embodiments of the present application are based on the following findings: Figure 2 It can be seen that the extracted sleep features are stable for the same individual, and are heterogeneous for different individuals.

[0106] The above-described embodiments only express several preferred embodiments of the present application, and the description is relatively specific and detailed, but it should not be understood as a limitation on the patent scope of the present application. It should be noted that for ordinary skilled in the art, several improvements and replacements can be made without departing from the technical principles of the present application, and these improvements and replacements should be considered as the protection scope of the present application. Therefore, the patent protection scope of the present application should be subject to the protection scope of the claims.

[0107] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present application, and not to limit them; although the present application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that they can still modify the technical solutions recorded in the foregoing embodiments, or make equivalent replacements for some or all of the technical features; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the scope of the technical solutions of the embodiments of the present application.

Claims

1. A sleep feature extraction method based on probability density function and polysomnography, characterized in that: The following steps are involved: Step S1), collecting and preprocessing the polysomnography (PSG) data of the subject, and labeling the sleep stages of the preprocessed PSG data; Step S2), dividing the pre-processed PSG data into time windows of fixed length, and extracting a set of time window feature groups for each time window in each polysomnography PSG channel, each time window feature group including time domain features, frequency domain features, and time-frequency domain features; Step S3), grouping the time window features in the time window feature group according to the sleep stage, and estimating the joint probability density function of the time window features on all PSG channels in each sleep stage; Step S4), using a contour map to visualize the joint probability density function; Step S5), obtaining comprehensive statistical characteristics of the polysomnography (PSG) based on the joint probability density function, and then predicting the age of the subject based on the comprehensive statistical characteristics of the polysomnography (PSG); The method for obtaining the comprehensive statistical characteristics of the polysomnography (PSG) in step S5) is as follows: Step S5.1) First, extract basic statistical features, including: The mean of the features of each time window in each channel within each sleep stage; Standard deviation of the features of each time window in each channel within each sleep stage; Correlation coefficients between time window features in each channel within each sleep stage; Step S5.2) Select non-rapid eye movement stage 2 N2 as the reference sleep stage and extract the intra-channel temporal features, including: In non-rapid eye movement phase 1 N1, non-rapid eye movement phase 3 N3 and rapid eye movement REM, in each channel, the difference between the mean of each time window feature and the mean of the time window feature in non-rapid eye movement phase 2 N2; In non-rapid eye movement phase 1 N1, non-rapid eye movement phase 3 N3 and rapid eye movement REM, in each channel, the standard deviation of each time window feature relative to the standard deviation of the time window feature in non-rapid eye movement phase 2 N2; The difference between the correlation coefficients of the time window features in each channel during non-rapid eye movement phase 1 N1, non-rapid eye movement phase 3 N3 and rapid eye movement REM relative to the correlation coefficient within non-rapid eye movement phase 2 N2; Step S5.3) For each PSG channel class containing multiple channels, a reference channel is selected from all the channels, and based on the selected reference channel, the inter-channel spatial features of this class of channels are extracted, including: In each sleep stage, in the channels other than the reference channel, the difference between the mean of each time window feature and the mean of the time window feature in the reference channel; In each sleep stage, in the channels other than the reference channel, the difference between the standard deviation of each time window feature and the standard deviation of the time window feature in the reference channel; In each sleep stage, the difference between the correlation coefficients between the time window features in the remaining channels except the reference channel and the correlation coefficients between the time window features in the reference channel is calculated.

2. The sleep feature extraction method based on probability density function and polysomnography according to claim 1, characterized in that: The polysomnography (PSG) data collected in step S1) contains multi-channel electrophysiological signals, including at least two electroencephalogram (EEG) channels, and the minimum sampling rate of each channel is 100 Hz. The polysomnography (PSG) data is preprocessed to reduce noise in the multi-channel electrophysiological signals.

3. The sleep feature extraction method based on probability density function and polysomnography according to claim 1, characterized in that: The specific method of labeling the sleep stages of the pre-processed PSG data in step S1) is as follows: according to the American Academy of Sleep Medicine AASM standard, the pre-processed PSG data is divided into 30s standard sleep frames, and each standard sleep frame is classified into five different sleep stages: Wake, non-rapid eye movement stage 1 N1, non-rapid eye movement stage 2 N2, non-rapid eye movement stage 3 N3 and rapid eye movement REM.

4. The sleep feature extraction method based on probability density function and polysomnography according to claim 3, characterized in that: The duration of the standard sleep frame is a multiple of the duration of the time window divided in step S2), so that each standard sleep frame includes one or more time windows.

5. The sleep feature extraction method based on probability density function and polysomnography according to claim 1, characterized in that: In step S2), different time window feature groups are extracted for time windows for different categories of polysomnography (PSG) channels. The function expression of the time window feature group is as follows: in, Indicates the mth channel c m The corresponding time window feature group; Indicates channel c m The kth time window feature under c m is the PSG channel identifier; m is the ordinal number of the channel; For channel c m The total number of corresponding time window features, k is the channel c m The ordinal number of the corresponding time window feature.

6. The sleep feature extraction method based on probability density function and polysomnography according to claim 1, characterized in that: In step S3), a joint probability density estimation is performed on the time window features of the four sleep stages of non-rapid eye movement (NREM) 1, non-rapid eye movement (NREM) 2, non-rapid eye movement (NREM) 3, and rapid eye movement (REM). The expression of the joint probability density function p(X) of the time window features on all PSG channels is as follows: Sd={N1,N2,N3,REM} Where p(X) represents the joint probability density function of the time window features of all PSG channels; s represents the sleep stage, Sd represents the set of the other four sleep stages except Wake; p s (X) represents the joint probability density function of the time window features under sleep stage s; P(s) represents the probability that the corresponding time window feature belongs to sleep stage s; Indicates channel c m The kth time window feature under ; M represents the total number of channels; X represents the time window feature set of all channels.

7. The sleep feature extraction method based on probability density function and polysomnography according to claim 1, characterized in that: The probability P(s) that the time window belongs to sleep stage s is obtained by the following equation: Where |s| represents the number of standard sleep frames in sleep stage s, and the denominator Σ s′∈Sd |s′| represents the total number of sleep frames.

8. The method for visualizing stable sleep characteristics based on probability density function according to claim 1, characterized in that: The step S4) is specifically as follows: Step S4.1) First, select a pair of time window feature pairs containing two different time window features. For sleep stage s, the time window feature joint probability density function p is obtained according to the following formula: s (X) About time window feature pairs The marginal distribution probability density function of Where X' represents the removal of All time window features except ; Indicates channel c p The i-th time window feature under ; Indicates channel c q The j-th time window feature under ; Step S4.2), draw the time window feature is the horizontal axis, time window characteristics The marginal distribution probability density function of the vertical axis Contour map, as the feature pair of this time window Contour plot of the marginal distribution probability density function under sleep stage s; Step S4.3) Repeat steps S4.1) to S4.2) until the time window feature pair is drawn in the same coordinate system. The marginal distribution probability density function contour map of all sleep stages in the corresponding sleep stage set Sd is then used to map the time window feature pairs. Perform visual display.

9. The sleep feature extraction method based on probability density function and polysomnography according to claim 1, characterized in that: The method of predicting the subject's age by comprehensive statistical features in step S5) is as follows: First, an age prediction dataset is constructed, including the comprehensive statistical characteristics of polysomnography (PSG) and its corresponding real age. Then, an age prediction model is constructed, and the age prediction dataset is input into the constructed age prediction model for training to obtain a trained age prediction model. Finally, the polysomnography (PSG) data of the test subject is input into the trained age prediction model to output the age prediction result of the test subject.

Images