CN-122004888-A - Semi-supervised learning obstructive sleep apnea OSA recognition method based on non-contact electrocardio

CN122004888ACN 122004888 ACN122004888 ACN 122004888ACN-122004888-A

Abstract

The application discloses a non-contact electrocardiosignal-oriented obstructive sleep apnea OSA recognition method which comprises the steps of (a) establishing and training a non-contact electrocardiosignal quality classification fusion model, carrying out denoising treatment on a 2-level non-contact electrocardiosignal obtained in the step (a), carrying out segmentation treatment on the 1-level and 2-level non-contact electrocardiosignals to obtain a plurality of sections of non-contact electrocardiosignals, (c) carrying out R peak recognition on each section of non-contact electrocardiosignal obtained in the step (b), extracting R wave peak values, RR intervals and RR interval first-order derivatives, thus obtaining input data comprising four channels corresponding to each section of non-contact electrocardiosignal, and (d) inputting tagged input data comprising four channels and non-tagged input data obtained in the step (c) into an OSA recognition deep learning model, and carrying out semi-supervised training on the recognition deep learning model through a Fixmatch algorithm. The method can automatically evaluate the quality of the noncontact ECG signal and simplify the detection process of OSA.

Inventors

WANG KAICHEN
CHEN CHEN
CHEN WEI
CUI RUI
ZHENG XIONGWEN
LI YANG

Assignees

粤港澳大湾区精准医学研究院(广州)
上海国际人类表型组研究院

Dates

Publication Date: 20260512
Application Date: 20241111

Claims (10)

1. The method for constructing the method for identifying the obstructive sleep apnea OSA based on the noncontact electrocardio is characterized by comprising the following steps of: (a) Establishing and training a non-contact electrocardiosignal quality classification fusion model, wherein the quality classification fusion model is configured to divide non-contact electrocardiosignal data into three classes, the three classes comprise a class 1, a class 2 and a class 3, the class 1 refers to a signal which can see that all clear R waves can be directly used for OSA identification, the class 2 refers to a signal which has a large amount of noise and cannot see that the R waves can be used for OSA identification after being processed, and the class 3 refers to a signal which has obvious motion artifact interference and cannot be used for OSA identification; (b) Denoising the 2-level non-contact electrocardiosignals obtained in the step (a), and then segmenting the 1-level and 2-level non-contact electrocardiosignals to obtain multi-segment non-contact electrocardiosignals; (c) Carrying out R peak identification on each section of non-contact electrocardiosignal obtained in the step (b), and extracting an R wave peak value, an RR interval and an RR interval first order derivative, so as to obtain input data which corresponds to each section of non-contact electrocardiosignal and comprises four channels, wherein the four channels respectively correspond to each section of non-contact electrocardiosignal, the R wave peak value of the section of non-contact electrocardiosignal, the RR interval of the section of non-contact electrocardiosignal and the RR interval first order derivative of the section of non-contact electrocardiosignal; (d) And inputting the tagged input data and the untagged input data containing the four channels into an OSA recognition deep learning model, and performing semi-supervised training on the OSA recognition deep learning model through Fixmatch algorithm, so as to obtain a prediction result of whether each section of non-contact electrocardiosignal contains OSA.
2. The method of claim 1, wherein the non-contact cardiac signal is segmented every 0.5-2 minutes in step (b).
3. The method of claim 1, wherein in step (a), the mass classification fusion model comprises a first residual neural network model and a two-way long-short-term memory network model, non-contact electrocardiosignal data with mass classification labels obtained by direct acquisition is input to the two-way long-term memory network model, and a time spectrum image of the non-contact electrocardiosignal data with mass classification labels is input to the first residual neural network model, so that training of the mass classification fusion model is achieved.
4. The method of claim 3, wherein the two-way long-short-term memory network model and the first residual neural network model are fused to form the quality classification fusion model by a decision mechanism, wherein the decision mechanism refers to taking an average value of output prediction vectors of the two-way long-short-term memory network model and the first residual neural network model, wherein the prediction vectors comprise probability values of each category, and a category with the largest probability value is taken as a prediction category of the non-contact electrocardiosignal data.
5. The method of claim 1, further comprising, prior to step (a), the step (a 0) of acquiring non-contact electrocardiographic signal data based on a capacitive coupling principle, performing segmentation processing on the acquired non-contact electrocardiographic signal data, and then adding a quality classification label to each segment of non-contact electrocardiographic signal data in combination with manual labeling, thereby forming training data of the quality classification fusion model.
6. The method of claim 1, wherein the OSA-recognition deep learning model is a convolutional neural network model CNNT comprising a second residual neural network model and a transducer encoder.
7. A method of identifying obstructive sleep apnea OSA, comprising: (a) Acquiring non-contact electrocardiosignal data acquired based on a capacitive coupling principle; (b) Inputting the non-contact electrocardiosignal data into a trained non-contact electrocardiosignal quality classification fusion model of any one of claims 1 to 6, wherein the non-contact electrocardiosignal data is divided into three classes, the three classes comprise a class 1, a class 2 and a class 3, the class 1 refers to a signal which can see that all clear R waves can be directly used for OSA identification, the class 2 refers to a signal which has a large amount of noise, the R waves cannot be seen to be used for OSA identification after being processed, and the class 3 refers to a signal which has obvious motion artifact interference and cannot be used for OSA identification; (c) Denoising the 2-level non-contact electrocardiosignals obtained in the step (b), and then segmenting the 1-level and 2-level non-contact electrocardiosignals to obtain a plurality of segments of non-contact electrocardiosignals; (d) Carrying out R peak identification on each section of non-contact electrocardiosignal obtained in the step (c), and extracting an R wave peak value, an RR interval and an RR interval first order derivative, so as to obtain input data which corresponds to each section of non-contact electrocardiosignal and comprises four channels, wherein the four channels respectively correspond to each section of non-contact electrocardiosignal, the R wave peak value of the section of non-contact electrocardiosignal, the RR interval of the section of non-contact electrocardiosignal and the RR interval first order derivative of the section of non-contact electrocardiosignal; (e) Inputting the input data comprising the four channels obtained in the step (d) into the OSA recognition deep learning model trained in any one of claims 1-6, thereby obtaining a prediction result of whether each section of non-contact electrocardiosignal contains OSA.
8. A device for non-contact electrocardiographic-based recognition of obstructive sleep apnea OSA, comprising: a memory for storing computer executable instructions, and A processor coupled to the memory for implementing the steps in the method of claim 7 when executing the computer-executable instructions.
9. A computer-readable storage medium comprising, the computer-readable storage medium has stored therein computer-executable instructions, the computer executable instructions, when executed by a processor, implement the steps in the method of claim 7.
10. A computer program product comprising computer-executable instructions which, when executed by a processor, implement the steps in the method of claim 7.

Description

Semi-supervised learning obstructive sleep apnea OSA recognition method based on non-contact electrocardio Technical Field The application relates to the field of sleep medicine in biomedicine, in particular to a semi-supervised learning obstructive sleep apnea OSA recognition method based on non-contact electrocardio. Background Good sleep is an important guarantee of normal life, obstructive sleep apnea (osthole SLEEP APNEA, OSA) is one of the common sleep disorders, which has adverse effects on both short-term and long-term health of humans, and is clinically manifested as snoring, somnolence, anxiety and irritability. In severe cases, it can lead to hypertension, stroke, and even heart-lung failure. Accurate clinical detection of OSA relies primarily on various overnight physiological signal recordings acquired by the PSG, such as ECG (Electrocardiograph, ECG), electroencephalogram (Electroencephalograph, EEG), arterial oxygen saturation (SaO 2), eye and leg movements, and respiration. The standard PSG data is mainly collected in a sleep disorder ward or a sleep laboratory of a hospital, so that the measurement cost is high, and a plurality of measurement electrodes cause inconvenience to a subject. At the same time, the physician may also experience errors due to fatigue in processing the multi-channel signals over a long period of time. Therefore, it is important to provide an acceptable monitoring modality and an accurate and simple diagnostic method for a large number of OSA patients. ECG is the most widely used diagnostic tool in cardiology. Understanding and interpreting ECG signals is critical to improving diagnostic accuracy and timely treatment of heart rhythm abnormalities, heart attacks, and OSA patients. The respiratory rhythm of the human body affects the sinus Fang Jiejie rhythm by way of the hertz-primary reflex, causing an associated change in the electrocardiogram with the respiratory rhythm. Thus, the ECG signal may provide effective diagnostic information in respiratory related diseases. The ECG signal of PSG is collected with wet electrodes attached to the skin, which would cause skin damage over time, and is not suitable for home long-term electrocardiographic monitoring. With the rapid development of sensing technology, the development of portable non-contact ECG monitoring has been greatly promoted. Non-contact measurement is an important tool for long-term monitoring of ECG due to the advantages of portability, no disturbance and low cost. However, non-contact measurement methods are sensitive to noise, including baseline wander, electromagnetic signals, and motion artifacts, among others. Poor signal quality can affect subsequent ECG signal analysis, making it difficult for a person to understand the correct diagnostic information. Therefore, it is important to find an efficient and accurate automatic signal quality assessment method for non-contact electrocardiography. Meanwhile, a rapid and accurate OSA recognition method is developed, the diagnosis efficiency of a patient is improved, the burden is reduced for doctors, errors caused by fatigue diagnosis of the doctors are avoided to the greatest extent, and the method is also a popular direction of OSA recognition research at present. In 2000 PhysioNet presented a challenge aimed at providing an inexpensive solution for detecting OSA using only a single lead ECG. However, the series of methods proposed before mostly require complicated manual feature extraction, feature selection and complex threshold determination, require a great deal of knowledge in the related field, do not achieve true automatic detection, and also bring task burden to doctors. Therefore, with the development of portable non-contact ECG monitoring devices and the popularization of daily health monitoring concepts, there is an urgent need in the art to develop a semi-supervised learning OSA identification method for non-contact electrocardio, which can automatically evaluate the quality of non-contact ECG signals and simplify the detection process of OSA, so that not only can the ECG signals monitored for a long time be better utilized, but also more people can know their sleep health status, and at the same time, the burden of medical staff is reduced. Disclosure of Invention The application aims to provide a semi-supervised learning OSA recognition method based on non-contact electrocardio, which can automatically evaluate the quality of non-contact ECG signals and simplify the detection process of OSA, so that the ECG signals monitored for a long time can be better utilized, more people can know the sleep health state of the people, and meanwhile, the burden of medical staff is reduced. The first aspect of the present application provides a method for constructing a method for identifying obstructive sleep apnea OSA based on noncontact electrocardio, comprising: (a) Establishing and training a non-contact electrocardiosignal quality classification fusi