Sleep posture classification method based on pseudo-item identification fusion and HMM algorithm
By non-invasively acquiring BCG signals and combining Group-Lasso feature filtering and pseudo-item identification with the HMM algorithm, the overfitting problem caused by pseudo-items in sleep posture recognition is solved, improving classification accuracy and reducing equipment costs.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- CHENGDU SANKESHAN INTELLIGENT MEDICAL TECH CO LTD
- Filing Date
- 2021-12-30
- Publication Date
- 2026-06-02
AI Technical Summary
Existing technologies for sleep posture recognition suffer from pseudo-terms that lead to model overfitting and low classification accuracy. Furthermore, traditional equipment has a highly intrusive data collection process, resulting in a poor user experience.
The study employed non-invasive BCG signal acquisition, performed preliminary feature extraction using HRV, filtered OM feature variables using Group-Lasso features, combined with pseudo-term identification and fusion methods, and finally performed sleep posture recognition based on the HMM algorithm.
It improves the accuracy of sleep posture classification, reduces equipment costs, provides a non-invasive data collection method, and enhances the user experience.
Smart Images
Figure CN114464317B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of sleep monitoring technology, and in particular to a sleep posture classification method based on pseudo-term identification fusion with OM feature variables and HMM algorithm. Background Technology
[0002] A Hidden Markov Model (HMM) is a statistical model used to describe a Markov process with hidden, unknown parameters. The challenge lies in determining these hidden parameters from the observable parameters. These parameters are then used for further analysis, such as pattern recognition. It is a statistical Markov model that models the system being modeled as a Markov process with unobserved (hidden) states.
[0003] Hidden Markov Models (HMMs) are a type of Markov chain. Their states cannot be directly observed, but can be observed through a sequence of observation vectors. Each observation vector represents various states through a certain probability density distribution, and each observation vector is generated by a sequence of states with a corresponding probability density distribution. Therefore, an HMM is a doubly stochastic process—a hidden Markov chain with a certain number of states and a set of explicit random functions. Since the 1980s, HMMs have been applied to speech recognition with great success. In the 1990s, HMMs were also introduced into computer character recognition and the core technology of "multi-user detection" in mobile communications. HMMs have also begun to be applied in fields such as bioinformatics and fault diagnosis.
[0004] In real life, ordinal multinomial (OM) variables are widely present. For example, the stages of tumor treatment can be divided into four levels, I-IV (another name for ordinal multinomial variables), representing the severity of a patient's condition from mild to severe. Based on the correlation of these level variables, researchers can examine other aspects of a patient's condition. However, in previous studies, the Hidden Markov Models (HMMs) typically neglected the preprocessing of OM variables, resulting in two or more adjacent categories of OM variables potentially having similar or even equivalent effects on the response variable. In other words, there are spurious terms among these adjacent categories, ultimately leading to overfitting of the model.
[0005] For example, patent application CN201310321043.5 discloses a sleep posture recognition method based on bioelectrical impedance analysis (BIA). This method involves collecting bioelectrical impedance signals from the left and right chest of the subject; performing analog-to-digital conversion on these signals to obtain digital bioelectrical impedance signals for the left and right chests respectively; extracting reference feature values based on these signals; and then using a hierarchical support vector machine (HSV) classification method to identify the subject's sleep posture based on these reference feature values. This method relies on bioelectrical impedance data for sleep posture recognition and classification. However, it requires the subject to wear measuring electrodes during data collection, resulting in a poor user experience and the data collection process being easily affected by the subject's posture. The classification accuracy is also low, requiring further improvement. Summary of the Invention
[0006] The purpose of this invention is to overcome the shortcomings of existing technologies and provide a sleep posture classification method based on pseudo-item recognition fusion and HMM algorithm. This method involves non-invasive data acquisition of BCG signals, preliminary feature extraction using HRV, filtering OM feature variables using Group-Lasso features, further processing the feature variables using pseudo-item recognition and fusion methods, and finally, sleep posture recognition based on the HMM algorithm. This method can provide data support for human sleep management and health monitoring through posture judgment, resolving the overfitting phenomenon caused by the presence of pseudo-items in the feature variables, thus overcoming the previous problem of low accuracy in sleep posture classification.
[0007] The objective of this invention is achieved through the following technical solution:
[0008] Sleep posture classification methods based on pseudo-item recognition fusion and HMM algorithm include:
[0009] Step 1: Signal preprocessing. Acquire the patient's raw BCG signal and perform preprocessing and feature extraction to obtain OM feature variables;
[0010] Step 2: Feature filtering and pseudo-item identification and fusion. Group-Lasso feature filtering is performed on the OM feature variables, and pseudo-items are identified and fused on the filtered OM feature variables to obtain the observation sequence.
[0011] Step 3: Sleep posture recognition. By continuously shifting the window width to extract the observation sequence as input to the Hidden Markov Model (HMM), the HMM is trained to obtain the final HMM. The actual observation sequence is then input into the model to identify the category of sleep posture.
[0012] Specifically, step one includes the following sub-steps:
[0013] S101: Data acquisition, using a mattress detection system to collect sleep data from the patient and obtain the patient's raw BCG signal;
[0014] S102: Normalize and preprocess the collected raw BCG signal data;
[0015] S103: Perform HRV signal analysis and feature extraction on the normalized preprocessed raw BCG signal data, and initially extract OM feature variables.
[0016] Furthermore, sub-step S102 specifically includes: firstly, using the maximum-minimum method, mapping the original BCG signal data to the interval [0, 1]; then, discretizing the mapped data with a segment length of 0.02 to obtain the original BCG signal data after normalization preprocessing.
[0017] Further, sub-step S103 specifically includes: performing time-domain analysis on the normalized preprocessed raw BCG signal data to obtain the variation of the RR interval of the raw BCG signal data over a certain period of time, and deriving a series of time-domain indicators, including: the standard deviation of the RR interval; the standard deviation of the mean RR interval for every 5-minute segment; the number and percentage of adjacent cardiac cycles with a difference greater than 50 milliseconds; the maximum, minimum, peak, kurtosis, first quantile, second quantile, third quantile, and fourth quantile of the RR interval, thereby initially obtaining characteristic variables; using the first quantile, second quantile, third quantile, and fourth quantile among the characteristic variables as OM characteristic variables, and the remaining characteristic variables as continuous variables.
[0018] Specifically, step two includes the following sub-steps:
[0019] S201: Encoding preprocessing, performing dummy variable encoding on the initially extracted OM feature variables;
[0020] S202: Variable selection, using the Group-Lasso feature selection method to select OM feature variables and continuous variables after dummy variable encoding;
[0021] S203: The first, second, third, and fourth quantiles of the variables containing OM feature variables are selected by the pseudo-item identification method and then fused to become new feature variables. The fusion process specifically includes: the first and second quantiles are fused into new feature variable Z1, and the third and fourth quantiles are fused into new feature variable Z2; and the fused feature variables Z1 and Z2 are used as the observation sequence.
[0022] Step three specifically includes: continuously shifting the window width to extract the observation sequence as input to the Hidden Markov Model, analyzing and training the Hidden Markov Model and outputting the final Hidden Markov Model; then inputting the actual observation sequence into the model for identification to obtain the specific category of the patient's sleep posture.
[0023] The beneficial effects of this invention are:
[0024] 1. This invention acquires BCG signals non-invasively, performs preliminary feature extraction using HRV, filters OM feature variables using Group-Lasso features, further processes feature variables using pseudo-term identification and fusion methods, and finally identifies sleeping postures based on the HMM algorithm, primarily recognizing supine, prone, left lateral, and right lateral positions. This posture determination provides data support for sleep management and health monitoring.
[0025] 2. This algorithm is the first to use a model that combines the pseudo-term identification and fusion method of OM feature variables with the HMM algorithm to classify sleep postures. This solves the problem of model overfitting caused by the presence of pseudo-terms in feature variables, thus overcoming the previous problem of low accuracy in sleep posture classification. Attached Figure Description
[0026] Figure 1 This is a flowchart of the method of the present invention;
[0027] Figure 2 This is a schematic diagram of the basic hidden Markov model of the present invention;
[0028] Figure 3 This is a graph showing the relationship between BCG signals and sleeping position;
[0029] Figure 4 This is a BCG signal graph when lying flat;
[0030] Figure 5 This is a flowchart of the model building and training process;
[0031] Figure 6 This is a flowchart of the model classification process. Detailed Implementation
[0032] To provide a clearer understanding of the technical features, objectives, and beneficial effects of this invention, the following detailed description of the technical solution is provided. Obviously, the described embodiments are only a portion of the embodiments of this invention, not all of them, and should not be construed as limiting the scope of implementation of this invention. Based on the embodiments of this invention, all other embodiments obtained by those skilled in the art without inventive effort are within the protection scope of this invention.
[0033] Example 1:
[0034] In this embodiment, as Figure 1 As shown, a sleep posture classification method based on pseudo-item recognition fusion and HMM algorithm includes:
[0035] Step 1: Signal preprocessing. Acquire the patient's raw BCG signal and perform preprocessing and feature extraction to obtain OM feature variables;
[0036] Step 2: Feature filtering and pseudo-item identification and fusion. Group-Lasso feature filtering is performed on the OM feature variables, and pseudo-items are identified and fused on the filtered OM feature variables to obtain the observation sequence.
[0037] Step 3: Sleep posture recognition. By continuously shifting the window width to extract the observation sequence as input to the Hidden Markov Model (HMM), the HMM is trained to obtain the final HMM. The actual observation sequence is then input into the model to identify the category of sleep posture.
[0038] This embodiment discloses a feature screening, pseudo-term identification, and fusion technique for OM (Ordinal Multinomial) feature variables. First, the Group-Lasso feature screening method is used to select variables for ordinal multinomial and continuous variables. Then, a pseudo-term identification method is combined with an HMM (Hidden Markov Model) model and applied to sleep posture recognition. Our patented technology detects and fuses OM feature variables in BCG signals, enabling better detection and identification of patients' sleep postures based on the original BCG signals. Typically, for individuals who are bedridden for extended periods, quantitative measurement and assessment of their sleep postures not only allows nurses to better assist them but also enables health monitoring systems to detect any hidden medical conditions.
[0039] Human posture lying position (HBLP) detection is crucial for adjusting patient posture. Generally, HBLP is divided into four categories: supine, prone, left lateral, and right lateral. Currently, sleep data collection via PSG is a dedicated device for monitoring sleep status and diagnosing sleep-related disorders. However, its high cost and the inconvenience it causes to the subjects are significant challenges. Therefore, this invention utilizes a non-invasive method to collect sleep data through mattress sensors. Features are extracted from the raw BCG signals, and Group-Lasso feature filtering is used to select OM (Occupational Sleep) feature variables. Pseudo-term identification technology is used to identify and fuse pseudo-terms within the OM feature variables. Finally, the Hidden Markov Model (HMM) algorithm is used to detect and classify sleep postures, identifying the patient's sleep position during sleep to help healthcare professionals better monitor patient conditions.
[0040] The sleep posture classification method in this embodiment is mainly implemented in the following four parts:
[0041] Part 1: Signal preprocessing, the specific steps are as follows:
[0042] S1. Data acquisition: Obtain the raw BCG signal through the mattress detection system;
[0043] S2. Perform normalization preprocessing on the collected data;
[0044] S3. Extract HRV features from the preprocessed data;
[0045] S4. Perform Group-Lasso feature filtering on the extracted OM feature variables;
[0046] S5. The pseudo-item identification technology identifies and fuses pseudo-items in the filtered OM feature variables;
[0047] S6. Input the processed feature variables into the HMM model for training and detect sleep posture categories.
[0048] Generally, the acquired raw BCG signal needs to be normalized first. The normalization method used in this algorithm is the minimum-maximum method, which maps the signal to the interval [0, 1] to eliminate modeling errors caused by individual differences. Then, the normalized data is discretized with a segment length of 0.02, which effectively reduces the time complexity. After that, HRV signal analysis is performed on the preprocessed signal to extract features. The Group-Lasso feature is used to filter OM feature variables. Then, pseudo-item recognition technology is used to identify and fuse the filtered features to obtain the fused feature variables as the observation sequence. The observation sequence is then continuously shifted by the window width and used as the input of HMM for HMM training to recognize sleeping postures.
[0049] In this embodiment, the mattress detection system uses a mattress-type multi-parameter health monitoring and recording system developed by Chengdu Lexiang Smart Home Technology Co., Ltd., which is used to collect raw BCG signals.
[0050] Typically, HRV signal analysis methods mainly include traditional time-domain analysis, frequency-domain analysis, and nonlinear analysis. Time-domain analysis is the earliest and most widely used method. It uses mathematical statistics to analyze the variation of RR intervals over a period of time, deriving a series of time-domain indicators. Currently, a set of statistical indicators from 1996 is commonly used as the standard, mainly including: the standard deviation of RR intervals; the standard deviation of the mean RR intervals every 5 minutes; the number and percentage of adjacent cardiac cycles with a difference greater than 50 milliseconds; and the maximum, minimum, peak, kurtosis, first quantile, second quantile, third quantile, and fourth quantile of RR intervals. This initially yields characteristic variables. In our method, since the first, second, third, and fourth quantiles are rank-based characteristic variables, they can be considered as OM characteristic variables in the Group-Lasso feature selection method. The remaining features are treated as continuous variables and incorporated into the pseudo-term identification and fusion method to obtain the final characteristic variables.
[0051] Part Two: Group-Lasso Feature Selection, Pseudo-item Recognition, and Fusion Algorithm
[0052] Step 1: Data preprocessing, dummy variable encoding for OM feature variables. In the feature processing stage of machine learning, categorical features are inevitably used. These features enter the model in a different way than general numerical variables.
[0053] Typically, depending on the needs of the model, categorical features need to be processed as dummy variables, that is, encoded according to the feature category. Generally, a feature with category k needs to be encoded as a set of k-1 derived dummy variables [to avoid multicollinearity]. This way, all categories within the feature can be represented (the baseline comparison class is set to 0, and when all k-1 dummy variables are 0, it is the baseline class).
[0054] Step 2: Variable selection. The Group-Lasso feature selection method is used to select OM feature variables and continuous variables.
[0055] Group-Lasso:
[0056]
[0057] Where ||·|| represents the norm. When p j When ||θ = 1, ||θ j ||2=|θ j |
[0058] Step 3: For simplicity, assume that after Group-Lasso filtering, there are q groups of OM feature variables and p continuous variables remaining.
[0059] Yi =α+β2X 2,i +...+β q X q,i +α1Z 1,i +...+α p Z p,i +ε i ,
[0060] Step 4: Substitute into the above equation and perform an equivalent transformation:
[0061]
[0062] Step 5: Organize:
[0063]
[0064] Step 6: Let
[0065]
[0066] Step 7: Use penalized regression to identify spurious classifications, let
[0067]
[0068] in and Two different penalty functions are used because the OM feature variables and continuous variables are fundamentally different.
[0069] This pseudo-item identification method can fuse the first, second, third, and fourth quantiles of the OM feature variables into new feature variables: the first and second quantiles are fused into a new feature variable Z1, and the third and fourth quantiles are fused into a new feature variable Z2.
[0070] The relevant parameters involved in the above Group-Lasso feature selection process are referenced from two papers: "Unbalanced Data Credit Scoring Model Based on Group-Lasso Method" and "Application Research of Generalized Semi-parameter Additive Credit Scoring Model Based on Group-LASSO Method".
[0071] Part 3: Hidden Markov Models (HMMs)
[0072] Hidden Markov Models (HMMs) are used to analyze and model the BCG signal after processing OM feature variables, thereby achieving sleep posture recognition. Figure 2 express, Figure 2The arrows in the diagram represent the relationships between the states. At any given time, an HMM describes a Markov process with hidden unknown parameters. Its hidden state X cannot be directly observed, but can be inferred from changes in the observation vector y. The establishment of a Hidden Markov Model mainly consists of the following three parameters:
[0073] (1) State transition probability matrix — A = {a ij} N×N
[0074] a ij =P(q) t+1 =S j |q t =S i ),i,j∈[1,N]
[0075] Let time t be from S i State transition at time t+1 S j The probability of a state.
[0076] (2) The probability matrix of observed values under given conditions — B = {b ij},
[0077] b ij =P(x t =c j |y t =S j ), i∈[1,N],j∈[1,M]
[0078] That is, the observed value generated in state j is c. j The probability of.
[0079] (3) Initial state probability distribution — π = {π i}
[0080] π i =P(y1=S) i ), i∈[1,N]
[0081] A Hidden Markov Model (HMM) can be represented as λ = (π, A, B). In practice, using an HMM requires solving three fundamental problems: First, the probability calculation problem: given the model λ = (π, A, B) and the observed values, the probability of the observed sequence occurring needs to be calculated. Second, the learning problem: given the observed values, the three parameters in the HMM model need to be determined to maximize the probability of the observed sequence. Third, the prediction problem: given the observed values and model parameters, the HMM model is used to predict the most likely output sequence.
[0082] Generally, the learning problem refers to estimating the parameters λ = (π, A, B) of the Hidden Markov Model (HMM) for each sleep posture using the maximum likelihood method. For parameter estimation, we choose the Baume-Welch algorithm. Finally, when calculating the sleep posture classification using measured data, the observed values are the processed data from the collected BCG signals; the goal is to calculate the corresponding sleep posture classification based on the observed signals.
[0083] Part 4: The Relationship Between BCG Signals and Sleep Posture
[0084] like Figure 3 and Figure 4 As shown, different sleeping positions exhibit different fluctuations in BCG signals. Therefore, we can select sleeping positions based on different characteristics of BCG signals for identification.
[0085] Example 2:
[0086] Based on the Hidden Markov Model constructed in Example 1, this example further presents the model construction and training process and the sleep posture classification process.
[0087] like Figure 5 As shown, the Hidden Markov Model (HMM) construction and training process includes: reading the original BCG signal and performing data preprocessing; extracting preliminary feature variables using HRV; filtering OM feature variables using Group-Lasso features; identifying and fusing feature pseudo-terms on the filtered feature variables; and using the processed feature variables as input to the HMM for training to obtain the finally trained HMM.
[0088] like Figure 6 As shown, the sleep posture classification process includes: reading the original BCG signal and performing data preprocessing, performing preliminary feature extraction through HRV, using Group-Lasso features to filter OM feature variables, further processing the feature variables using pseudo-term identification and fusion methods, and finally performing sleep posture recognition based on the trained Hidden Markov Model to output the sleep posture recognition result.
[0089] This invention acquires BCG signals non-invasively, performs preliminary feature extraction using HRV, filters OM feature variables using Group-Lasso features, further processes feature variables using pseudo-term identification and fusion methods, and finally identifies sleeping postures based on the HMM algorithm, primarily recognizing supine, prone, left lateral, and right lateral positions. This posture determination provides data support for sleep management and health monitoring.
[0090] This invention's HMM algorithm is the first to combine the OM feature variable pseudo-term identification and fusion method with the HMM algorithm to classify sleep postures. This solves the problem of model overfitting caused by the presence of pseudo-terms in the feature variables, thereby overcoming the previous problem of low accuracy in sleep posture classification.
[0091] Previously, PSG (Pressure Sensor) devices were used for signal acquisition, but these devices were too expensive. This embodiment uses a mattress detection system for BCG (Body Shape Detection) signal acquisition, significantly reducing costs. Furthermore, the model for sleep posture recognition, which combines pseudo-term identification and fusion methods of OM (Optical Model) feature variables with the HMM (Hypertext Markov Model) algorithm, greatly improves the model's accuracy.
[0092] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely illustrative of the principles of the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the present invention as claimed. The scope of protection of this invention is defined by the appended claims and their equivalents.
Claims
1. A sleep posture classification method based on pseudo-item recognition fusion and HMM algorithm, characterized in that, include: Step 1: Signal preprocessing. Acquire the patient's raw BCG signal and perform preprocessing and feature extraction to obtain OM feature variables; Step 2: Feature filtering and pseudo-item identification and fusion. Group-Lasso feature filtering is performed on the OM feature variables, and pseudo-items are identified and fused on the filtered OM feature variables to obtain the observation sequence. Step 3: Sleep posture recognition. The observation sequence is continuously shifted and truncated by the window width as the input of the Hidden Markov Model (HMM) for training. The final HMM is obtained and the actual observation sequence is input into the model to identify the category of sleep posture. Step one specifically includes the following sub-steps: S101: Data acquisition, using a mattress detection system to collect sleep data from the patient and obtain the patient's raw BCG signal; S102: Normalize and preprocess the collected raw BCG signal data; S103: Perform HRV signal analysis and feature extraction on the normalized preprocessed raw BCG signal data, and initially extract OM feature variables; Step two specifically includes the following sub-steps: S201: Encoding preprocessing, performing dummy variable encoding on the initially extracted OM feature variables; S202: Variable selection, using the Group-Lasso feature selection method to select OM feature variables and continuous variables after dummy variable encoding; S203: The first, second, third, and fourth quantiles of variables containing OM feature variables are selected by pseudo-term identification method and then fused to become new feature variables. The fusion process specifically includes: the first and second quantiles are fused into new feature variable Z1, and the third and fourth quantiles are fused into new feature variable Z2; and the fused feature variables Z1 and Z2 are used as the observation sequence. Step three specifically includes: continuously shifting the window width to extract the observation sequence as input to the Hidden Markov Model, analyzing and training the Hidden Markov Model and outputting the final Hidden Markov Model; then inputting the actual observation sequence into the model for recognition to obtain the category of the patient's sleep posture.
2. The sleep posture classification method based on pseudo-item recognition fusion and HMM algorithm according to claim 1, characterized in that, Sub-step S102 specifically includes: first, using the maximum and minimum value method, mapping the original BCG signal data to the interval [0, 1]; then, discretizing the mapped data with a segment length of 0.02 to obtain the original BCG signal data after normalization preprocessing.
3. The sleep posture classification method based on pseudo-item recognition fusion and HMM algorithm according to claim 1, characterized in that, Sub-step S103 specifically includes: performing time-domain analysis on the normalized preprocessed raw BCG signal data to obtain the variation of the RR interval of the raw BCG signal data over a certain period of time, and obtaining a series of time-domain indicators, including: the standard deviation of the RR interval; the standard deviation of the mean RR interval for every 5-minute segment; the number and percentage of adjacent cardiac cycles with a difference greater than 50 milliseconds; the maximum, minimum, peak, kurtosis, first quantile, second quantile, third quantile, and fourth quantile of the RR interval, thereby initially obtaining characteristic variables; using the first quantile, second quantile, third quantile, and fourth quantile of the characteristic variables as OM characteristic variables, and the remaining characteristic variables as continuous variables.