Model training method and device, sleep monitoring method and device and electronic equipment

Abstract

The invention provides a model training method and device, a sleep monitoring method and device and electronic equipment. A training method of a sleep monitoring model based on transfer learning, the method comprising: acquiring a polysleep monitoring (PSG) dataset and a ballistocardiogram (BCG) dataset, the PSG dataset at least comprising a PSG respiratory signal and a PSG heartbeat signal, and the BCG dataset at least comprising a BCG respiratory signal and a BCG heartbeat signal; pre-training the sleep monitoring initial model based on the PSG data set to obtain a sleep monitoring base model; and performing fine tuning on the sleep monitoring base model based on the BCG data set to obtain a sleep monitoring model.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of sleep monitoring, and in particular, to a model training method and device for sleep monitoring, a sleep monitoring method and device, an electronic device, a computer device and a storage medium. BACKGROUND

[0002] Non-invasive and accurate sleep monitoring has become an important research focus because both sleep deficiency and sleep abnormalities are associated with a range of health problems. These problems include daily issues (such as weight gain and impaired driving performance), as well as specific sleep disorders such as sleep apnea, and other conditions (such as Parkinson's disease, depression, and Alzheimer's disease). Effective sleep monitoring helps to diagnose sleep disorders and identify early intervention for related health problems.

[0003] However, current clinical sleep diagnosis relies on polysomnography (PSG), the gold standard for sleep assessment. PSG provides comprehensive analysis by simultaneously recording multiple physiological signals, including electroencephalogram (EEG), electrocardiogram (ECG), electrooculogram (EOG), electromyogram (EMG), and airflow, etc. Although PSG has high precision, PSG has significant drawbacks: PSG is very costly, with a very high price; PSG has a very poor user experience because it requires multiple electrodes and sensors to be attached to the user, making it very inconvenient for the user to sleep, and PSG is complex to construct, generally requiring the user to go to a hospital to use it; PSG has poor continuous monitoring capability, and from a cost perspective, users often use PSG to sleep for only 1-2 nights for sleep monitoring, which cannot achieve continuous sleep monitoring. Therefore, balancing comfort, cost, and diagnostic performance remains a key challenge for developing alternative sleep monitoring devices.

[0004] To address these limitations, a portable sleep monitoring system using ballistocardiogram (BCG) has been designed. This system provides a non-invasive and cost-effective alternative to capture mechanical movements caused by heart activity. However, the BCG signal is derived from body micro-movements, and any body movement, turning over, speaking, or clothing friction will superimpose noise, resulting in a decrease in signal-to-noise ratio and a significant decrease in recognition accuracy. In addition, most studies on sleep monitoring using BCG have only been validated in healthy adults or small samples, lacking large-scale, multi-disease (children, the elderly, sleep disorder patients) clinical data support. Therefore, due to its insufficient data volume and sensitivity to motion noise, sleep monitoring based on BCG still faces challenges, including limited detection accuracy and poor generalization. SUMMARY

[0005] To solve the above technical problems, the application provides a sleep monitoring method based on transfer learning. Specifically, it uses large-scale PSG data as a source domain to guide the learning process on small-sample BCG data in a target domain. In this method, the model is first pre-trained on PSG records using two channels of ECG-derived heartbeats and abdomen (ABD)-derived respiration. This pre-training stage enables the network to acquire a robust sleep diagnosis representation. In the subsequent fine-tuning stage, these learned representations are transferred to BCG data based on transfer learning by preserving the corresponding heartbeat and respiration paths. This strategy allows the model to benefit from the diagnostic knowledge extracted from large-scale PSG signals, ultimately improving the performance of sleep staging and apnea-hypopnea index (AHI) evaluation.

[0006] According to an aspect of the application, a model training method based on transfer learning for sleep monitoring is provided, which includes: obtaining a polysomnography PSG data set and a ballistocardiogram BCG data set, wherein the PSG data set includes at least PSG respiration signals and PSG heartbeat signals, and the BCG data set includes at least BCG respiration signals and BCG heartbeat signals; pre-training a sleep monitoring initial model based on the PSG data set to obtain a sleep monitoring base model; and fine-tuning the sleep monitoring base model based on the BCG data set to obtain a sleep monitoring model.

[0007] According to an embodiment of the application, the sleep monitoring model is used to at least perform a sleep staging task and / or a sleep apnea estimation task.

[0008] According to an embodiment of the application, pre-training the sleep monitoring initial model based on the PSG data set includes training using the PSG respiration signals and the PSG heartbeat signals as a training set.

[0009] According to an embodiment of the application, fine-tuning the sleep monitoring base model based on the BCG data set includes fine-tuning the sleep monitoring base model using the BCG respiration signals and the BCG heartbeat signals, and additional bandpass signals, wherein the additional bandpass signals include bandpass filtered BCG respiration signals and bandpass filtered BCG heartbeat signals.

[0010] According to an embodiment of the application, the additional bandpass signals further include a bandpass filtered BCG snoring signal.

[0011] According to an embodiment of the present application, the sleep monitoring model can include a CNN feature extraction module, an adaptive feature fusion module, and a Roformer encoding module, and wherein: the CNN feature extraction module is configured to perform feature extraction on the BCG respiratory signal and the BCG heartbeat signal, and the band-pass filtered BCG respiratory signal, the band-pass filtered BCG heartbeat signal, and the band-pass filtered BCG snoring signal, respectively, to obtain enhanced features of the BCG respiratory signal and the BCG heartbeat signal, and band-pass features of the band-pass filtered BCG respiratory signal, the band-pass filtered BCG heartbeat signal, and the band-pass filtered BCG snoring signal; the adaptive feature fusion module is configured to perform feature fusion on the enhanced features and the band-pass features to obtain fused features; and the Roformer encoding module is configured to process the fused features to perform sleep staging or sleep apnea estimation.

[0012] According to an embodiment of the present application, the sleep monitoring model further includes an Adam optimizer configured to dynamically adjust the learning rate for the sleep staging task and the sleep apnea estimation task.

[0013] According to an embodiment of the present application, the sleep monitoring model outputs the sleep apnea severity by estimating the AHI value, and the sleep apnea severity includes mild, moderate, and severe.

[0014] According to an embodiment of the present application, the PSG heartbeat signal includes an RRI heartbeat sequence, and the BCG heartbeat signal includes a JJI heartbeat sequence.

[0015] According to an embodiment of the present application, the method further includes performing signal quality assessment on the PSG dataset and the BCG dataset, and selecting the PSG respiratory signal and the PSG heartbeat signal for pre-training and the BCG respiratory signal and the BCG heartbeat signal for fine-tuning based on the assessment results.

[0016] According to an embodiment of the present application, the signal quality assessment includes: calculating a confidence score, which is the ratio between the sum of the energy of the PSG or BCG respiratory signal and the PSG or BCG heartbeat signal in each time period and the total energy in the time period; and calculating the average confidence score over all time periods, wherein, in the case that the average confidence score is greater than a threshold, the PSG or BCG respiratory signal and the PSG or BCG heartbeat signal are retained for pre-training or fine-tuning of the sleep monitoring base model, and wherein, in the case that the average confidence score is less than or equal to the threshold, the PSG or BCG respiratory signal and the PSG or BCG heartbeat signal are discarded.

[0017] According to an embodiment of the present application, the threshold is 0.6.

[0018] According to an embodiment of the present application, the PSG dataset is obtained from a plurality of sources, the plurality of sources comprising one or more of SHHS, MESA, MROS, CHAT, WSC, PATS, NUMOM2B, and HSP.

[0019] According to another aspect of the present application, a sleep monitoring method is provided, wherein the method comprises: collecting a BCG respiratory signal and a BCG heartbeat signal of a user; and analyzing the BCG respiratory signal and the BCG heartbeat signal using a sleep monitoring model obtained by the training method of the sleep monitoring model based on transfer learning to obtain a sleep staging result and an AHI severity for the user.

[0020] According to yet another aspect of the present application, a training apparatus of a sleep monitoring model based on transfer learning is provided, wherein the apparatus comprises: a component for obtaining a polysomnography (PSG) dataset and a ballistocardiogram (BCG) dataset, wherein the PSG dataset comprises at least a PSG respiratory signal and a PSG heartbeat signal, and the BCG dataset comprises at least a BCG respiratory signal and a BCG heartbeat signal; a component for pre-training an initial sleep monitoring model based on the PSG dataset to obtain a sleep monitoring base model; and a component for fine-tuning the sleep monitoring base model based on the BCG dataset to obtain a sleep monitoring model.

[0021] According to another aspect of the present application, a sleep monitoring apparatus is provided, wherein the apparatus comprises: a component for collecting a BCG respiratory signal and a BCG heartbeat signal of a user; and a component for analyzing the BCG respiratory signal and the BCG heartbeat signal using a sleep monitoring model obtained by the training method of the sleep monitoring model based on transfer learning to obtain a sleep staging result and an AHI severity for the user.

[0022] According to another aspect of the present application, an electronic device is provided, the electronic device comprising the sleep monitoring apparatus.

[0023] According to another aspect of the present application, a computer device is provided, comprising a memory and a processor, wherein the memory stores computer readable instructions, and the computer readable instructions, when executed by the processor, implement the training method of the sleep monitoring model based on transfer learning and the sleep monitoring method.

[0024] According to another aspect of the present application, a computer readable storage medium is provided, wherein the computer readable storage medium stores computer readable instructions, and the computer readable instructions, when executed by a processor, implement the training method of the sleep monitoring model based on transfer learning and the sleep monitoring method.

[0025] Therefore, the training method and device of the sleep monitoring model based on transfer learning, the sleep monitoring method and device, the electronic device, the computer device and the storage medium provided in the embodiments of the present application can perform sleep staging and sleep apnea estimation only by using BCG data, thereby eliminating the need for PSG. The above-mentioned technical solution has low cost and reliable diagnostic performance, is simple to operate for users, and can perform sleep monitoring in a non-invasive form under comfortable conditions. BRIEF DESCRIPTION OF DRAWINGS

[0026] In order to more clearly illustrate the technical solutions of the embodiments of the present application, the drawings needed in the description of the embodiments will be briefly introduced. Obviously, the drawings in the following description are only some example embodiments of the present application, and other drawings can be obtained by those skilled in the art without creative labor on the basis of these drawings.

[0027] Figure 1 A first flowchart of the training method of the sleep monitoring model based on transfer learning according to the embodiments of the present application is shown;

[0028] Figure 2 A source distribution diagram of the PSG data set according to the embodiments of the present application is shown;

[0029] Figure 3 A second flowchart of the training method of the sleep monitoring model based on transfer learning according to the embodiments of the present application is shown;

[0030] Figure 4 A schematic diagram of the improved fine-tuning stage in the training method of the sleep monitoring model based on transfer learning according to the embodiments of the present application is shown;

[0031] Figure 5 A flowchart of the sleep monitoring method according to the embodiments of the present application is shown;

[0032] Figure 6 A comparison diagram of the test results of the sleep monitoring method according to the embodiments of the present application and the traditional sleep monitoring method is shown;

[0033] Figure 7 A schematic diagram of multi-center BCG queue verification according to the embodiments of the present application is shown;

[0034] Figure 8 A structural schematic diagram of the training device of the sleep monitoring model based on transfer learning according to the embodiments of the present application is shown;

[0035] Figure 9 A structural schematic diagram of the sleep monitoring device according to the embodiments of the present application is shown;

[0036] Figure 10 Fig. 6 shows a schematic diagram of a result displayed on a terminal device according to an embodiment of the present application and output by a sleep monitoring device according to an embodiment of the present application; and

[0037] Figure 11 Fig. 7 shows a structural schematic diagram of a computer device according to an embodiment of the present application. DETAILED DESCRIPTION

[0038] In order to make the objectives, technical solutions and advantages of the embodiments of the present application clearer, the technical solutions of the embodiments of the present application will be described clearly and completely below with reference to the drawings of the embodiments of the present application. Obviously, the described embodiments are only some of the embodiments of the present application, rather than all the embodiments of the present application. Based on the described embodiments of the present application, all other embodiments obtained by those of ordinary skill in the art without creative effort belong to the scope of protection of the present application.

[0039] Unless otherwise defined, technical terms or scientific terms used in the present application shall have the common meaning understood by one of ordinary skill in the art to which the present application pertains. The terms “first”, “second”, and similar terms used in the present application do not denote any order, quantity, or importance, but are used to distinguish different components. The terms “include”, “contain”, and similar terms mean that the elements or objects before the terms encompass the elements or objects listed after the terms and their equivalents, and do not exclude other elements or objects. The terms “connect” or “connected” and similar terms do not mean only physical or mechanical connections, but can include electrical connections, whether direct or indirect. The terms “up”, “down”, “left”, “right”, and the like only represent relative positional relationships, and when the absolute positions of the described objects are changed, the relative positional relationships can also be changed accordingly. In order to keep the following description of the embodiments of the present application clear and concise, the present application omits the detailed description of some known functions and known components.

[0040] Flowcharts are used in the present application to illustrate the steps of the methods according to the embodiments of the present application. It should be understood that the preceding or subsequent steps do not necessarily proceed in order. On the contrary, various steps can be processed in reverse order or simultaneously. Meanwhile, other operations can be added to these processes, or a step or several steps can be removed from these processes.

[0041] In the specification and drawings of the present application, elements are described in singular or plural form according to the embodiments. However, the selection of singular and plural forms for the proposed case is only for the convenience of explanation and does not intend to limit the present application to this. Therefore, the singular form can include the plural form, and the plural form can also include the singular form, unless the context clearly indicates otherwise.

[0042] In order to better understand the scheme of the embodiments of the present application, the related terms and concepts that may be involved in the embodiments of the present application are introduced first. It should be understood that the explanation of the related concepts may be limited to the specific situation of the embodiments of the present application, but it does not mean that the present application can only be limited to this specific situation, and the specific situation of different embodiments may also be different, and the specific situation is not limited here.

[0043] PSG: a medical examination method for evaluating the sleep quality of patients and diagnosing sleep disorders, which is considered as the "gold standard" for diagnosing sleep disorders. PSG can record multiple physiological parameters simultaneously, including electroencephalogram (EEG), electrooculogram (EOG), electromyogram (EMG), electrocardiogram (ECG), respiratory airflow, respiratory effort, blood oxygen saturation (SpO2), snoring sound, body position, and limb movement.

[0044] BCG: a mixed signal based on heartbeat, which integrates multiple physiological signals. Generally, BCG is a weak vibration signal caused by the change of external pressure of human body due to the heart beat and the blood circulation of aorta. Since the chest rises and falls and the diaphragm moves during breathing, which produces mechanical action on the heart and surrounding tissues, it also causes regular vibration of the body, which affects the presentation of BCG signal, so that BCG presents a "sinusoidal-like wave" signal with larger amplitude and lower frequency. At this time, the heartbeat signal is the superimposed peak vibration signal with smaller amplitude and higher frequency. Similarly, the heart condition, blood vessel condition and posture change of human body also more or less make BCG present different regular signal output. Therefore, by analyzing the strength and waveform change of BCG signal, the basic state of the heart condition, blood vessel condition, breathing condition and posture change of human body can be judged.

[0045] ECG: electrocardiogram, ECG signal is an electrical signal, in each cardiac cycle of the heart, the pacemaker, atrium, ventricle are excited in turn, accompanied by the change of bioelectricity, through electrocardiograph, the graph of potential change in multiple forms is drawn from the body surface. Among them, the current in the myocardium is generated before the myocardial contraction during the contraction of the heart, and the change of the myocardial current is ECG.

[0046] RRI: the waveform is a complete ECG signal, wherein the solid line represents a heart beat waveform with complete wave band components, including P wave, PR wave band, QRS wave group, S-T wave band and T wave. For ECG signal, the time interval between adjacent two R waves is RR interval (R-R interval, RRI), which represents the rhythm information of heart beat. After obtaining a certain length of ECG signal, the RRI in each complete ECG signal can be extracted, so that the RRI sequence of ECG signal can be obtained.

[0047] JJI: The BCG signal waveform can be divided into ballistic waves F, G, H, I, J, K, L, M, N according to different stages. There are 6 complete BCGs in the BCG signal waveform, among which there are 8 ballistic waves from G to N in some BCGs, and only 6 ballistic waves from G to L in some BCGs. For the BCG signal, the time interval between adjacent two J peaks is simply referred to as "J-J interval (JJI)". After obtaining a certain length of BCG signal, the JJI in each complete BCG signal can be extracted, so that the JJI sequence of the obtained BCG signal can be obtained.

[0048] AHI: is an important indicator for assessing the severity of sleep apnea. It measures the severity of the disease by calculating the number of apnea and hypopnea events per hour of sleep. The adult AHI evaluation standard is: normal: AHI < 5 times / hour; mild: 5≤AHI<15 times / hour; moderate: 15≤AHI<30 times / hour; severe: AHI≥30 times / hour.

[0049] CNN: is a deep neural network with convolutional structure, which is a deep learning architecture. Deep learning architecture refers to the algorithm of machine learning that learns at multiple levels on different levels of abstraction. As a deep learning architecture, CNN is a feed-forward artificial neural network, in which each neuron responds to overlapping regions in the input image.

[0050] ResNet: is a CNN with "residual modification". ResNet solves the problems of gradient disappearance and gradient explosion in the training of deep neural networks by introducing a residual learning mechanism based on classical CNN, making it possible to train very deep networks.

[0051] Encoder: is responsible for converting input sequences into high-dimensional feature representations. The encoder usually contains multiple self-attention layers and feed-forward neural network layers. Each self-attention layer uses a self-attention mechanism to assign a weight to each position in the input sequence, and then calculates the weighted sum value based on these weights, thereby capturing long-distance dependencies in the input sequence.

[0052] ResNet Encoder: is an encoder module based on the ResNet (Residual Network) architecture, widely used in computer vision and deep learning tasks.

[0053] RoFormer: A Transformer model combined with Rotary Position Encoding (RoPE) to more effectively integrate position information to improve the model's ability to process long texts.

[0054] The neural network model mentioned in the present application can include various types, such as deep neural networks (DNN), convolutional neural networks (CNN), recurrent neural networks (RNN), generative adversarial networks (GAN), Transformer, long short-term memory networks (LSTM), residual networks (ResNet), or other neural networks, etc., and can also be a combination of the above-mentioned various neural network models, and the present application does not limit the type of neural network model.

[0055] BCG has emerged as a non-invasive alternative to PSG, which captures the overall performance of the circulatory system by measuring mass movement. During atrial contraction, blood jets into large blood vessels, shifting the body's center of mass toward the head. Conversely, when blood moves toward peripheral blood vessels, the center of mass shifts toward the feet. This change is influenced by cardiac activity, respiration, and body movement, producing BCG waveforms as blood distribution changes throughout the cardiac cycle.

[0056] The differences in data distribution between PSG and BCG mainly come from differences in data acquisition methods, signal characteristics, and subject-specific factors (e.g., age, health status). PSG directly measures high-resolution, multi-modal physiological signals (e.g., EEG, EMG, EOG, and ECG), which provide precise information about sleep stages and brain activity. In contrast, BCG captures the mechanical vibrations of the body produced by cardiac and respiratory activity, resulting in lower-resolution signals that are more susceptible to noise and motion artifacts.

[0057] These intrinsic differences lead to mismatches between the feature spaces and statistical distributions of PSG and BCG data. Notably, BCG lacks direct measurements of brain activity and exhibits higher noise levels, leading to distribution shifts that pose significant challenges when attempting to transfer knowledge from models developed on PSG data to BCG data. Addressing this transition is crucial for developing robust, cross-modal algorithms for sleep monitoring.

[0058] The model training process for sleep monitoring provided by the embodiments of the present application will be described in detail below with reference to the accompanying drawings.

[0059] Figure 1A first flowchart S100 of a training method of a sleep monitoring model based on transfer learning according to embodiments of the present application is shown. The various steps of the training method of the sleep monitoring model based on transfer learning according to embodiments of the present application will be described in detail below with reference to Figure 1

[0060] First, as shown in Figure 1 The first flowchart S100 of the training method of the sleep monitoring model based on transfer learning according to embodiments of the present application can include the various steps shown according to processes S1020, S1040 and S1060.

[0061] In process S1020, a polysomnography (PSG) dataset and a ballistocardiogram (BCG) dataset can be obtained, where the PSG dataset can include at least a PSG respiration signal and a PSG heartbeat signal, and the BCG dataset can include at least a BCG respiration signal and a BCG heartbeat signal.

[0062] In some embodiments, publicly available PSG datasets are considered to be collected from multiple large-scale sleep studies to ensure diversity and robustness of the model training. The PSG datasets include:

[0063] • Sleep Heart Health Study (SHHS): A multicenter cohort study investigating the impact of sleep-disordered breathing on cardiovascular health. It includes overnight PSG recordings of more than 6,000 participants (mainly middle-aged and older adults).

[0064] • Cleveland Children's Sleep and Health Study (CCSHS): A study that collects sleep data of children, including sleep duration, sleep quality, symptoms of sleep disorders, and the relationship between these sleep characteristics and children's health. The data is mainly collected at the Cleveland University Hospital Medical Center.

[0065] CCSHS mainly collects data at the Cleveland University Hospital Medical Center. This study collects sleep data of children, including sleep duration, sleep quality, symptoms of sleep disorders, and the relationship between these sleep characteristics and children's health.

[0066] • Children's tonsillectomy study: A randomized clinical trial examining the impact of adenotonsillectomy on children with obstructive sleep apnea. The dataset includes overnight PSG recordings of 464 children aged 5-9.

[0067] • Home PAP study: A study that evaluates home and laboratory diagnosis and treatment methods for sleep apnea and provides overnight PSG and portable monitoring data.

[0068] • Multi-Ethnic Study of Atherosclerosis (MESA): A cohort study investigating the relationship between sleep characteristics and cardiovascular health, including PSG data from more than 2,000 adults of different ethnic backgrounds.

[0069] ​• Male Osteoporotic Fracture Study (MrOS): a longitudinal study exploring sleep disorders in older men (>65 years), PSG data from over 3,000 participants.

[0070] • Wisconsin Sleep Cohort (WSC): a longitudinal epidemiological study investigating the natural history of sleep apnea, characterized by the collection of overnight PSG data every four years from a cohort of middle-aged adults.

[0071] • Study of Osteoporotic Fractures (SOF): a large cohort study examining sleep disorders in older women, including PSG and actigraphy data.

[0072] • Pediatric Adenotonsillectomy Treatment for Snoring (PATS): a study evaluating the efficacy of adenotonsillectomy in children with habitual snoring but no severe obstructive sleep apnea (OSA), including detailed PSG assessment.

[0073] • Nulliparous Pregnancy Outcomes Study: Monitoring Mothers (NuMoM2B): a study evaluating sleep-disordered breathing in pregnant women and its impact on maternal and fetal health, using overnight PSG recordings from over 3,000 participants.

[0074] • Human Sleep Project (HSP): this dataset is an expanding compilation of clinical PSG recordings. Starting with PSG recordings from approximately 19,000 patients evaluated at Massachusetts General Hospital, HSP is set to expand over the coming years to encompass data from over 200,000 patients and individuals evaluated beyond the clinical setting.

[0075] In addition to the aforementioned PSG public datasets, other PSG datasets can be introduced to further augment the training data, thereby enhancing the diversity and robustness of the model training.

[0076] As Figure 2 shown, in one example, a large-scale PSG dataset is collected from multiple sources, and incomplete labels and corrupted records are filtered out.

[0077] In process S1040, a sleep monitoring initial model can be pre-trained based on the PSG dataset to obtain a sleep monitoring base model.

[0078] In embodiments of the present application, the specific type of training model for sleep monitoring is not limited, and any one of the aforementioned neural network models or combinations thereof can be used as the base model in embodiments of the present application, as long as it can achieve the above-mentioned sleep monitoring function after training.

[0079] In some examples, pre-training the sleep monitoring initial model based on the PSG dataset can include training using the PSG respiration signal and the PSG heartbeat signal as a training set.

[0080] During the pre-training process, the filtered PSG recordings are used as the source domain, which is used to train the deep learning model. The model consists of several layers of CNN feature extractors and a transformer-based backbone. The CNN performs convolutional computation to extract local and compressed features from the input signals. These extracted features are then concatenated to form long-range temporal features, which are subsequently fed into several layers of transformer networks for global information processing. At this stage, the deep learning model acquires sleep-related knowledge from PSG data.

[0081] Specifically, during the pre-training stage, the sleep monitoring initial model is pre-trained using two channels from PSG recordings to obtain a sleep monitoring base model: heartbeat (ECG-based RRI) and respiration (ABD belt). Each channel is fed into a dedicated 1D ResNet class feature extractor, which consists of: 1) a convolutional layer (kernel size = 7, padding = 3), followed by batch normalization and ReLU activation; 2) a max-pooling layer (kernel size = 3, padding = 1); and 3) four consecutive residual blocks (each containing two ResBlock1D units). The extracted features from the two channels are summed element-wise to form a combined feature. A learnable CLS token is pre-added to this tensor, and the resulting sequence is fed into a RoFormer transformer encoder with an embedding size of 512. The final fully connected layer produces logits for sleep staging or AHI estimation. This stage learns a robust representation of heartbeat and respiration signals from the PSG domain.

[0082] In some examples, pre-training the sleep monitoring initial model based on the PSG dataset can further include testing the sleep monitoring base model's prediction ability of the electrical activity of the brain or heart using electrocardiogram signals or oximetry signals obtained from the PSG dataset as a test set.

[0083] Typically, when sleep monitoring is performed in a hospital using PSG, PSG will collect PSG electrocardiogram signals or PSG oximetry signals of the user in addition to PSG respiration signals and PSG heartbeat signals of the user. Therefore, after pre-training the sleep monitoring initial model using PSG respiration signals and PSG heartbeat signals as a training set, the PSG electrocardiogram signals or PSG oximetry signals predicted by the model can be compared with the collected real PSG electrocardiogram signals or PSG oximetry signals to test the prediction ability of the model for the electrical activity of the brain or heart.

[0084] In process S1060, the sleep monitoring base model can be fine-tuned based on the BCG dataset to obtain a sleep monitoring model.

[0085] In some examples, the method can further include performing signal quality assessment on the PSG dataset and the BCG dataset, and selecting the PSG respiratory signal and the PSG heartbeat signal for pre-training and the BCG respiratory signal and the BCG heartbeat signal for fine-tuning based on the assessment results.

[0086] In some examples, to ensure the reliability of PSG and BCG records used in model training, a signal quality assessment process is implemented based on the energy distribution within the respiratory and heartbeat frequency bands. Specifically, for each time segment, the ratio between the sum of the energy of the PSG or BCG respiratory signal and the PSG or BCG heartbeat signal and the total energy within the time segment is calculated. For each record, the average confidence score over all time segments is calculated. In the case where the average confidence score is greater than a threshold value, the PSG or BCG respiratory signal and the PSG or BCG heartbeat signal are retained for pre-training or fine-tuning of the sleep monitoring base model, and in the case where the average confidence score is less than or equal to the threshold value, the PSG or BCG respiratory signal and the PSG or BCG heartbeat signal are discarded. In some examples, the threshold value can be set to 0.6.

[0087] Through the above quality assessment process, it is ensured that only records with sufficiently high overall signal quality are included in the dataset. By maintaining comparable quality control strategies on both PSG and BCG modalities, the robustness of the model is improved and the introduction of low-quality physiological signals that can negatively affect training and evaluation is prevented.

[0088] Then, the original signals are processed through a respiratory signal extraction, a heartbeat signal extraction, and a body motion removal process.

[0089] Respiratory signal extraction: The respiratory signal includes both the signal obtained through specific signal processing on the original BCG signal and the ABD signal in PSG. First, a third-order Bessel bandpass filter (0.1-0.35 Hz) is implemented on the original BCG and ABD signals. The filtered signals are resampled to 4 Hz. For the BCG-derived respiratory signal, a time integration is performed to address the phase difference between BCG (indicating respiratory flux) and ABD (reflecting air volume). For the PSG-derived signal, the integration is omitted. Finally, both signals are standardized to z-score to eliminate the amplitude difference inherent to the sensor principle.

[0090] Heartbeat signal extraction: Heartbeat intervals are denoted as interbeat intervals (IBI). The PSG heartbeat signal comprises a sequence of RRIs, and the BCG heartbeat signal comprises a sequence of JJIs. Specifically, when the heartbeat signal is an ECG signal, the corresponding sequence of RRIs can be directly extracted from the ECG signal; when the heartbeat signal is a BCG signal, the corresponding sequence of JJIs can be pre-extracted from the BCG signal, and the corresponding sequence of RRIs can be obtained based on the mapping relationship between RRI and JJI.

[0091] Body motion detection: Body movements cause rapid fluctuations in the BCG and EEG signals. Given the 500 Hz sampling rate of these signals, a 2-second window is used to capture transient changes. The peak-to-peak amplitude within each window is used as a mobility feature. A dynamic 30-second baseline (divided into 15 consecutive 2-second windows) is established to compute the mean (μ) and standard deviation (σ) of the feature sequence. A threshold of μ+5σ is used to identify significant deviations, which minimizes false positives by excluding physiological fluctuations. This threshold aligns with the statistical norm of Gaussian-distributed data, ensuring robust detection of motion-induced artifacts.

[0092] In the fine-tuning phase, the sleep monitoring base model is trained on the target domain using BCG data to obtain a sleep monitoring model. Compared with the pre-training phase, the network structure remains unchanged. However, only the last few layers of the model are unfrozen for training on BCG data. This process, commonly referred to as "fine-tuning" in transfer learning, enables the model to retain knowledge from the source domain while adapting and learning new representations from the target domain.

[0093] Through this multi-source transfer learning framework, a model is obtained that can perform sleep staging and sleep apnea estimation using only BCG data as input. Unlike PSG, which requires multiple electrodes and sensors attached to the body, the method proposed in this application enables comfortable and completely contactless sleep monitoring while the patient is in bed. Sleep staging visualization and analysis results are ultimately transmitted to the user terminal.

[0094] Figure 3 A second flowchart S300 of a training method of a sleep monitoring model based on transfer learning according to an embodiment of the present application is shown.

[0095] As Figure 3 shown, the second flowchart S300 of the training method of the sleep monitoring model based on transfer learning can include various steps shown according to processes S3020, S3040 and S3060. Figure 3 Steps S3020 and S3040 in Figure 1 are the same as or similar to steps S1020 and S1040 in , and will not be described here again.

[0096] In Figure 3In the process S3060, the BCG respiratory signal and the BCG heartbeat signal, and the additional bandpass signals including the bandpass filtered BCG respiratory signal and the bandpass filtered BCG heartbeat signal can be used to fine-tune the sleep monitoring base model.

[0097] In addition, the additional bandpass signals can further include a bandpass filtered BCG snoring signal.

[0098] In combination Figure 4 In the improved fine-tuning stage of the training method of the sleep monitoring model based on the transfer learning according to the embodiments of the present application shown in the schematic diagram, the sleep monitoring model can include a CNN feature extraction module, an adaptive feature fusion module, and a Roformer encoding module. The CNN feature extraction module is configured to perform feature extraction on the BCG respiratory signal and the BCG heartbeat signal, and the bandpass filtered BCG respiratory signal, the bandpass filtered BCG heartbeat signal, and the bandpass filtered BCG snoring signal, respectively, to obtain the enhanced features of the BCG respiratory signal and the BCG heartbeat signal, and the bandpass features of the bandpass filtered BCG respiratory signal, the bandpass filtered BCG heartbeat signal, and the bandpass filtered BCG snoring signal; the adaptive feature fusion module is configured to perform feature fusion on the enhanced features and the bandpass features to obtain the fusion features; and the Roformer encoding module processes the fusion features to perform sleep staging or sleep apnea estimation.

[0099] In the improved fine-tuning stage, five channels are used to fine-tune the model on the BCG data: two transferred from the pre-training stage (BCG heartbeat and filtered respiratory) and three additional bandpass filtered BCG signals (snoring, heartbeat, respiratory).

[0100] Specifically, the network receives five input channels:

[0101] 1. BCG heartbeat (JJI) - transferred from the PSG heartbeat branch;

[0102] 2. BCG respiratory - transferred from the PSG respiratory branch;

[0103] 3. Bandpass filtered BCG snoring (0.5-25 Hz, sampling rate = 50 Hz);

[0104] 4. Bandpass filtered BCG heartbeat (30-200 Hz, sampling rate = 400 Hz);

[0105] 5. Bandpass filtered BCG respiratory (0.1-1 Hz, sampling rate = 2 Hz).

[0106] The first two channels utilize the weights learned from PSG (pre-training phase) to transfer the learned representations of heartbeat and respiration to the BCG domain. The three additional bandpass signals (snore, heartbeat, respiration) are each passed through a separate 1D ResNet feature extractor, whose architecture is similar to the pre-training phase extractor but with token length adapted to their respective sampling rates. A learnable gating mechanism is applied to these three bandpass features, enabling the model to adaptively weight their contributions. These gated bandpass features are then aggregated (element-wise addition) with the two transferred channels to form the fused feature representation. The fused representation is passed through the same RoFormer encoder and final fully connected layer after the preposed CLS token. This approach allows the model to leverage the knowledge transferred from the PSG channels and the additional discriminative information present in the bandpass filtered BCG signals.

[0107] In the improved fine-tuning phase, in addition to preserving the respective BCG heartbeat and respiration paths, the three bandpass filtered BCG channels (snore, heartbeat, and respiration) are additionally incorporated to convert these learned representations to BCG data. This dual-branch strategy allows the model to benefit from the diagnostic knowledge extracted from PSG signals and simultaneously capture unique features. Specifically, the BCG heartbeat signal and the BCG respiration signal extracted from the BCG heartbeat and respiration paths have discrete temporal features, while the bandpass filtered BCG respiration signal, the bandpass filtered BCG heartbeat signal, and the bandpass filtered BCG snore signal obtained via the bandpass filtered BCG channels (snore, heartbeat, and respiration) can be regarded as continuous raw signals (with some irrelevant features such as noise filtered out), which do not have significant statistical features and have more details. Therefore, by adopting the BCG heartbeat and respiration paths and the three bandpass filtered BCG channels (snore, heartbeat, and respiration), the significant statistical features and the raw signal features of the BCG signals can be combined, thereby further optimizing the sleep monitoring model.

[0108] In the present application, the sleep monitoring model is at least used to perform a sleep staging task and / or a sleep apnea estimation task. Table 1 summarizes the key training hyperparameters for the two tasks.

[0109] Table 1: Key training hyperparameters for the two tasks

[0110]

[0111] Specifically, the initial learning rate for both tasks is set to 1 × 10 −4 during fine-tuning. −5Cross-entropy loss was used for binary classification of sleep staging and apnea-hypopnea events, where apnea-hypopnea events were grouped into one class and normal respiratory events formed another class. The batch size was set to 32 in all experiments, and the data was split into training, validation, and test sets using an 8:1:1 ratio.

[0112] In some embodiments, the training set is mainly used to train the initial model; the validation set is mainly used to evaluate the effect of the model, and adjust the hyperparameters to put the model in the best state; the test set is mainly used to evaluate the performance of the model through a series of evaluation indexes such as accuracy, recall rate, F1, etc. Among them, the validation set is not necessary, if the hyperparameters do not need to be adjusted, the validation set can not be used, and the test set is directly used to evaluate the effect.

[0113] Sleep staging settings:

[0114] The duration of each input segment is 30s (token length = 30s). The whole night recording is divided into consecutive 30s segments for training and evaluation.

[0115] AHI estimation settings:

[0116] For AHI estimation, the data is segmented at 1s intervals (token length = 1s) and trained on 30-minute segments (input length = 30 minutes).

[0117] Sleep staging is crucial for assessing sleep quality and has been shown to influence a wide range of health conditions, such as cardiovascular disease, metabolic disorders, and neurodegenerative diseases. In this application, a four-class sleep stage classification is adopted following recent approaches (i.e., wake, N1+N2, N3, REM), where N1 and N2 are combined due to low consistency for N1. It should be noted that the division of sleep stages in the embodiments of the present application is not limited.

[0118] Sleep apnea syndrome, particularly obstructive sleep apnea (OSA), is a prevalent disorder characterized by recurrent episodes of partial or complete upper airway obstruction during sleep, leading to intermittent hypoxemia and fragmented sleep. This condition is clinically significant as it is associated with hypertension, stroke, and other serious comorbidities. The apnea-hypopnea index (AHI) is defined as the number of apnea and hypopnea events per hour of sleep. The severity of sleep apnea is generally classified as mild (5 < AHI ≤ 15), moderate (15 < AHI ≤ 30), or severe (AHI > 30). AHI is used as a key quantitative indicator for evaluating sleep apnea in this paper.

[0119] Figure 5 A flowchart of a sleep monitoring method according to an embodiment of the present application is shown. The following will be described with reference toFigure 5 The various steps of the sleep monitoring method according to the embodiments of the present application will be described in detail.

[0120] First, as shown in FIG. 5, the flowchart S500 of the sleep monitoring method according to the embodiments of the present application can include various steps according to processes S5020 and S5040. Figure 5

[0121] In the process S5020, the BCG respiratory signal and the BCG heartbeat signal of the user can be collected.

[0122] In some examples, the way of collecting the BCG respiratory signal and the heartbeat signal of the user can include air cushion monitoring. By air cushion monitoring, the respiration, heartbeat and body movement of the human body are measured by a pressure sensor, so as to identify various respiratory disorders.

[0123] In the process S5040, the BCG respiratory signal and the BCG heartbeat signal can be analyzed by using the sleep monitoring model obtained by the model training method as shown in FIG. 6 or FIG. 7, so as to obtain the sleep staging result and the AHI severity for the user. Figure 1 Figure 3

[0124] Figure 6 A comparison chart of test results of the sleep monitoring method according to the embodiments of the present application and the conventional sleep monitoring method is shown.

[0125] In some examples, the multi-source transfer learning (MSTL) method is compared with three baseline methods:

[0126] 1. PSG test based on PSG: This method is used as a high-performance standard for sleep staging, because PSG is considered as the gold standard for sleep monitoring due to its comprehensive physiological data and high reliability.

[0127] 2. BCG test based on BCG: This method simulates the deployment of sleep staging using only BCG data, representing the baseline performance of BCG in sleep monitoring.

[0128] 3. BCG zero-shot test based on PSG: This evaluates the performance of the PSG trained model when directly applied to BCG data without fine-tuning, testing its universality across modalities.

[0129] ​​​In some embodiments, a comprehensive set of performance metrics are employed to evaluate the proposed models, including Cohen's kappa, accuracy, macro F1 score, and weighted macro F1 score. Cohen's kappa coefficient KAPPA is a key metric in analysis that takes into account the possibility of coincidental agreement, making it a reliable indicator of inter-rater reliability. The values of KAPPA are interpreted as follows: KAPPA > 0.4 indicates moderate agreement, KAPPA > 0.6 indicates substantial agreement, and KAPPA > 0.8 indicates almost perfect agreement. Accuracy (ACC) is a simple metric calculated as the ratio of correctly predicted labels to the total number of segments, each of which has a label. Macro F1 score (MACRO F1) provides a balanced measure of precision and recall by averaging the F1 score for each class, regardless of its frequency. This makes it particularly useful in imbalanced datasets. Weighted F1 score (WEIGHTED F1) adjusts the macro F1 calculation by weighting the contribution of each class according to the number of 30-second data segments it contains. In other embodiments, the evaluation can also be performed in other ways, and the application does not limit the types of evaluation.

[0130] As shown in Figure 6 , the PSG-based PSG test achieves a higher level of performance, significantly outperforming the BCG-based BCG test and the PSG-based BCG zero-shot test. Notably, it achieves comparable performance to the proposed MSTL-based BCG test, with Cohen's kappa exceeding 0.76 for both methods and accuracy exceeding 0.85. Among the three baselines, the BCG-based BCG test exhibits the lowest performance, and this limitation is further amplified by class imbalance (accuracy: 0.743 → weighted F1: 0.736). The PSG-based BCG zero-shot test exhibits better performance than the BCG-based BCG test (accuracy: 0.815 vs. 0.746). However, when compared to the MSTL-based BCG test that includes transfer learning, a significant performance drop is observed. Specifically, Cohen's kappa decreases by 0.064 (0.78 → 0.716), and weighted F1 decreases by 0.045 (0.856 → 0.811), highlighting the effectiveness of MSTL in mitigating cross-domain discrepancies and enhancing model adaptation.

[0131] For the evaluation of AHI estimation, the F1 score is used, which provides a balanced measure of precision and recall, which is particularly important considering the potential class imbalance in the AHI estimation task. As shown in Figure 6As shown, the PSG-based PSG test achieves a high F1 score of 0.96, demonstrating robust performance in AHI estimation. Close behind is the MSTL-based BCG test, which reaches an F1 score of 0.962, slightly better than the PSG-based PSG test. This highlights the effectiveness of the MSTL framework in leveraging transfer learning to enhance model adaptability and performance in cross-domain scenarios. The BCG-based BCG test also performs well, reaching an F1 score of 0.958, which is almost comparable to the PSG-based PSG test. However, the PSG-based BCG zero-shot test shows a significant drop in performance, with an F1 score of 0.88. This drop underscores the challenges associated with zero-shot learning in cross-domain settings, where the lack of domain-specific fine-tuning can lead to reduced model accuracy and generalization.

[0132] Overall, the proposed MSTL-based BCG test achieves the highest performance among all methods, with a Cohen's kappa of 0.78, an accuracy of 0.856, and a weighted F1 score of 0.856.

[0133] By comparing the PSG-based PSG test and the BCG-based BCG test, the PSG-based approach significantly outperforms the BCG-based approach, as evidenced by an 11% gap in Cohen's kappa. This highlights the superiority of PSG in sleep staging due to its more rich physiological signals, including ECG and abdominal movements used.

[0134] According to the scheme proposed in the present application, it provides direct insights into sleep states and sleep transitions. In contrast, BCG relies on indirect signals derived from body movements and heart activity, making it inherently less detailed and more susceptible to noise.

[0135] Although PSG-based sleep staging achieves strong performance, its applicability is limited to PSG-specific settings. When directly applied to BCG inference, i.e., the PSG-based BCG zero-shot test, a significant performance drop is observed, with a 5.3% reduction in Cohen's kappa. This drop can be attributed to the domain disparity between PSG and BCG data distributions, which makes it challenging for a model trained only on PSG to effectively generalize to BCG.

[0136] The MSTL-based BCG test combines the fine-tuning through transfer learning, resulting in a significant performance boost over the PSG-based BCG zero-shot test (0.716→ 0.78 in Cohen's kappa). Notably, it even slightly outperforms the PSG-based PSG test (0.78 vs. 0.769 in Cohen's kappa). This enhancement can be attributed to two key advantages of transfer learning: i) knowledge diversity: by integrating multiple PSG datasets (source domains), the model learns a richer set of features and patterns, which enrich its knowledge to the BCG data (target domain); ii) improved generalization: including different PSG datasets provides a more comprehensive representation of sleep-related features, enhancing the adaptability to BCG data. These benefits are particularly evident in the classification of specific sleep stages. For example, as shown in Figure 6 the recall rate for N3 sleep stage improves from 0.526 to 0.707, marking a gain of nearly 30% over the PSG-based BCG test. Moreover, the F1 scores for all sleep stages confirm that the MSTL-based BCG test consistently outperforms other methods.

[0137] In summary, the MSTL-based BCG test is the best-performing method for sleep staging and AHI estimation, slightly outperforming the PSG-based PSG test, while the PSG-based BCG zero-shot test lags behind due to its inherent limitations in handling cross-domain discrepancies. These results further emphasize the importance of introducing transfer learning strategies, such as MSTL, to improve model robustness and accuracy in sleep monitoring.

[0138] Figure 7 A schematic diagram of a multi-center BCG cohort validation according to embodiments of the present application is shown.

[0139] As Figure 7 shown, to ensure the robustness and universality of the proposed method, a multi-center BCG cohort study was conducted in four hospitals: Shanghai Sixth People's Hospital (SSPH), West China Hospital (WCH), Suzhou University Affiliated Second Hospital (SAHSU), and Inner Mongolia Hospital (IMBM). Figure 7 The sleep staging performance and sleep apnea estimation performance of the above-mentioned sleep monitoring method proposed by the present application are shown respectively in the four hospitals. This multi-center validation framework not only confirms the consistency of the results across different clinical environments, but also highlights the adaptability of the method to different patient populations and data collection schemes.

[0140] Figure 8 A structural schematic diagram of a training device 800 of a sleep monitoring model based on transfer learning according to embodiments of the present application is shown.

[0141] Referring to Figure 8As shown, the apparatus 800 can include: a component 8020 for obtaining a polysomnography, PSG, data set and a ballistocardiogram, BCG, data set, wherein the PSG data set comprises at least a PSG respiratory signal and a PSG heartbeat signal, and the BCG data set comprises at least a BCG respiratory signal and a BCG heartbeat signal; a component 8040 for pre-training a sleep monitoring initial model based on the PSG data set to obtain a sleep monitoring base model; and a component 8060 for fine-tuning the sleep monitoring base model based on the BCG data set to obtain a sleep monitoring model, wherein the BCG data set comprises the BCG respiratory signal and the BCG heartbeat signal extracted from the BCG.

[0142] In a possible implementation, the component 8040 for pre-training can include: a component for training using the PSG respiratory signal and the PSG heartbeat signal as a training set.

[0143] In a possible implementation, the sleep monitoring model can be used at least for performing a sleep staging task and / or a sleep apnea estimation task.

[0144] In a possible implementation, the pre-training of the sleep monitoring initial model based on the PSG data set comprises: training using the PSG respiratory signal and the PSG heartbeat signal as a training set.

[0145] In a possible implementation, the fine-tuning of the sleep monitoring base model based on the BCG data set can include: fine-tuning the sleep monitoring base model using the BCG respiratory signal and the BCG heartbeat signal, and additional bandpass signals, wherein the additional bandpass signals comprise a bandpass filtered BCG respiratory signal and a bandpass filtered BCG heartbeat signal.

[0146] In a possible implementation, the additional bandpass signals can further comprise a bandpass filtered BCG snoring signal.

[0147] In a possible implementation, the sleep monitoring model can include a CNN feature extraction module, an adaptive feature fusion module, and a Roformer encoding module, and wherein: the CNN feature extraction module is configured to perform feature extraction on the BCG respiratory signal and the BCG heartbeat signal, and the bandpass filtered BCG respiratory signal, the bandpass filtered BCG heartbeat signal, and the bandpass filtered BCG snoring signal, respectively, to obtain enhanced features of the BCG respiratory signal and the BCG heartbeat signal, and bandpass features of the bandpass filtered BCG respiratory signal, the bandpass filtered BCG heartbeat signal, and the bandpass filtered BCG snoring signal; the adaptive feature fusion module is configured to perform feature fusion on the enhanced features and the bandpass features to obtain fused features; and the Roformer encoding module is configured to process the fused features to perform sleep staging or sleep apnea estimation.

[0148] In a possible implementation, the sleep monitoring model can further include an Adam optimizer configured to dynamically adjust a learning rate for the sleep staging task and the sleep apnea estimation task.

[0149] In a possible implementation, the sleep monitoring model can output the sleep apnea severity by estimating an AHI value, the sleep apnea severity including mild, moderate, and severe.

[0150] In a possible implementation, the PSG heartbeat signal can include an RRI heartbeat sequence, and the BCG heartbeat signal can include a JJI heartbeat sequence.

[0151] In a possible implementation, the apparatus 900 can further include a component configured to perform signal quality assessment on the PSG dataset and the BCG dataset, and select the PSG respiratory signal and the PSG heartbeat signal for pre-training and the BCG respiratory signal and the BCG heartbeat signal for fine-tuning based on the assessment result.

[0152] In a possible implementation, the signal quality assessment can include: calculating a confidence score, the confidence score being a ratio between a sum of energies of the PSG or BCG respiratory signal and the PSG or BCG heartbeat signal in each time period and a total energy in the time period; and calculating an average confidence score over all time periods, wherein, in a case where the average confidence score is greater than a threshold, the PSG or BCG respiratory signal and the PSG or BCG heartbeat signal are retained for pre-training or fine-tuning of the sleep monitoring base model, and wherein, in a case where the average confidence score is less than or equal to the threshold, the PSG or BCG respiratory signal and the PSG or BCG heartbeat signal are discarded.

[0153] In a possible implementation, the threshold can be set to 0.6.

[0154] In a possible implementation, the PSG dataset can be obtained from a plurality of sources, the plurality of sources can include one or more of SHHS, MESA, MROS, CHAT, WSC, PATS, NUMOM2B, and HSP.

[0155] Figure 9 A structural schematic diagram of a sleep monitoring apparatus according to an embodiment of the present application is shown.

[0156] Reference is made to Figure 9 As shown, the sleep monitoring apparatus 900 can include a component 9020 configured to collect a BCG respiratory signal and a BCG heartbeat signal of a user, and a component 9021 configured to use the BCG respiratory signal and the BCG heartbeat signal to perform pre-training of a sleep monitoring base model. Figure 1 or Figure 3The sleep monitoring model obtained by the illustrated model training method analyzes the BCG respiratory signal and the BCG heartbeat signal to obtain the sleep staging result and the AHI severity component for the user 9040.

[0157] Figure 10 A schematic diagram of a result output by the sleep monitoring apparatus according to an embodiment of the present application is shown.

[0158] As Figure 10 shown, the analysis result of sleep staging and the respiratory health estimation are transmitted to the user terminal in a visualized manner. The user can conveniently view the deep sleep state every day, the sleep condition of each time period, and the respiratory health condition, etc.

[0159] Figure 11 A structural diagram of a computer device according to an embodiment of the present application is shown.

[0160] Referring to Figure 11 , the computer device 110 can include a processor 1120 and a memory 1140. The processor 1120 and the memory 1140 can be connected through a bus 1130. The computer device 110 can be any type of portable device (such as a smart camera, a smart phone, a tablet computer, etc.) or any type of fixed device (such as a desktop computer, a server, etc.).

[0161] The processor 1120 can perform various actions and processes according to the programs stored in the memory 1140. Specifically, the processor 1120 can be an integrated circuit chip with signal processing capability. The above-mentioned processor can be a general-purpose processor, a digital signal processor (DSP), an application-specific integrated circuit (ASIC), a ready programmable gate array (FPGA) or other programmable logic device, a discrete gate or transistor logic device, a discrete hardware component. The disclosed methods, steps and logic block diagrams in the embodiments of the present application can be implemented or executed. The general-purpose processor can be a microprocessor or the processor can also be any conventional processor, etc., which can be X86 architecture or ARM architecture.

[0162] The memory 1140 stores computer-executable instructions that, when executed by the processor 1120, implement the training method and the sleep monitoring method of the sleep monitoring model based on transfer learning described above. The memory 1140 can be a volatile memory or a non-volatile memory, or can include both volatile and non-volatile memory. The non-volatile memory can be a read-only memory (ROM), a programmable read-only memory (PROM), an erasable programmable read-only memory (EPROM), an electrically erasable programmable read-only memory (EEPROM), or a flash memory. The volatile memory can be a random access memory (RAM) used as an external cache. By way of example, and not limitation, many forms of RAM can be used, such as static random access memory (SRAM), dynamic random access memory (DRAM), synchronous dynamic random access memory (SDRAM), double-data rate synchronous dynamic random access memory (DDR SDRAM), enhanced synchronous dynamic random access memory (ESDRAM), synchronous link dynamic random access memory (SLDRAM), and direct memory bus random access memory (DRAM). It should be noted that the memory of the methods described herein is intended to include, but not be limited to, these and any other suitable types of memory.

[0163] Embodiments of the present application also provide an electronic device, which can include Figure 9 The sleep monitoring device provided by the embodiments shown. In an embodiment, the electronic device can be a sleep staging detection device or a sleep state detection device. In an embodiment, the electronic device can be a non-contact detection device, for example, a smart air cushion or a smart headrest. In an embodiment, the electronic device can have an internal sensor and obtain user-related sensor data through the internal sensor, wherein the internal sensor can be a piezoelectric sensor, a piezoresistive sensor, or an optical fiber sensor. In other embodiments, the sensor in the electronic device for obtaining user-related sensor data can also be other types of sensors, which are not limited by the present application. In another embodiment, the electronic device can also obtain user-related sensor data such as respiration signals and heartbeat signals provided by external sensors through wireless or wired means.

[0164] In some embodiments, the electronic device obtains the user's breathing signal and heartbeat signal in a contact measurement and a non-contact measurement, wherein the contact measurement can include measurement by wearing a breathing belt, a bracelet, a watch, and the like. The non-contact measurement can include air cushion monitoring and microwave radar monitoring. The air cushion monitoring uses a pressure sensor to measure the breathing, heartbeat, and body movement of a human body, thereby identifying various respiratory disorders. This method is a non-contact method, comfortable for the tester, and suitable for home use by the user, but has a small range of action and its reliability is affected by temperature and humidity. In some embodiments, in the air cushion monitoring scenario, the electronic device can be a smart headrest or a smart air cushion, and the smart headrest or the smart air cushion can be internally provided with a piezoelectric sensor, a piezoresistive sensor, or an optical fiber sensor.

[0165] The microwave radar breathing monitoring method can use the micro-movement of the chest and abdomen during breathing to act on the radar echo signal, and extract the breathing signal by analyzing the radar transmission signal and the echo signal, with high accuracy. When used, the microwave radar device is placed at a certain vertical height from the human body at the head of the bed or the bedside, without contact with the human body, and can penetrate through quilts, clothes, and the like, with a large range of action and millimeter-level measurement accuracy. The microwave radar method has better measurement accuracy and accuracy than the mattress type breathing detection method while maintaining comfort, and is the preferred method for non-contact breathing measurement.

[0166] In addition, the training method of the sleep monitoring model based on transfer learning and the sleep monitoring method according to the present application can be stored in a computer readable storage medium. Specifically, according to the present application, a computer readable storage medium storing computer readable instructions can be provided, and the computer readable instructions, when executed by a processor, can cause the processor to execute the training method of the sleep monitoring model based on transfer learning and the sleep monitoring method as described above.

[0167] Embodiments of the present application also provide a computer program product including computer programs or instructions, which, when executed by a processor, implement the training method of the sleep monitoring model based on transfer learning and the sleep monitoring method provided by any of the embodiments of the present application.

[0168] It should be noted that the terminal involved in the embodiments of the present application can include, but is not limited to, a personal computer (PC), a personal digital assistant (PDA), a wireless handheld device, a tablet computer, a mobile phone, an MP3 player, an MP4 player, and the like.

[0169] It should be noted that the flowcharts and block diagrams in the drawings are illustrations of the possible architectures, functional processes, and operations of systems, methods, and computer program products in accordance with various embodiments of the present application. In this regard, each block in the flowcharts or block diagrams can represent a module, a segment, or a portion of code, which comprises one or more executable instructions for implementing the specified logical functions. It should also be noted that in some alternative implementations, the functions noted in the blocks can occur out of the order noted in the figures. For example, two blocks shown in succession may, in fact, be executed substantially concurrently or the blocks may sometimes be executed in the reverse order, depending upon the functionality involved. It will also be noted that each block of the block diagrams and / or flowchart illustrations, and combinations thereof, can be implemented by a dedicated hardware-based system that performs the specified functions or operations, or combinations of hardware and software.

[0170] In general, the various example embodiments of the application can be implemented in hardware or special-purpose circuits, software, firmware, logic, or any combination thereof. Some aspects of the application can be implemented in hardware, while other aspects can be implemented in firmware or software which can be executed by a controller, microprocessor or other computing device, although the application is not limited thereto. While various aspects of the application can be illustrated and described as block diagrams, flow charts, or using some other pictorial representation, it is well understood that these blocks, apparatus, systems, techniques or methods described herein can be implemented in, as non-limiting examples, hardware, software, firmware, special purpose circuits or logic, general purpose hardware or controler or other computing devices, or some combination thereof.

[0171] Unless otherwise defined, all terms (including technical and scientific terms) used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application belongs. It will be further understood that terms, such as those defined in commonly used dictionaries, should be interpreted as having a meaning that is consistent with their meaning in the context of the relevant art and will not be interpreted in an idealized or overly formal sense unless expressly so defined herein.

[0172] The foregoing is a summary of the present application and is not to be considered as limiting its scope. While several example embodiments of the application have been described, it will be apparent to those of ordinary skill in the art that many modifications are possible without departing from the novel teachings and advantages of the application. The embodiments chosen and described are meant to be illustrative only and are not intended to limit the scope of the application. It is to be understood that the above description is intended to be illustrative and not restrictive. Many embodiments will be apparent to those of skill in the art upon reviewing the above description. The scope of the application should, therefore, be determined not with reference to the above description, but should instead be determined with reference to the appended claims, along with their full scope of equivalents.

Claims

1. A method for training a sleep monitoring model based on transfer learning, the method comprising: obtaining a polysomnography (PSG) dataset and a ballistocardiogram (BCG) dataset, wherein the PSG dataset comprises at least a PSG respiratory signal and a PSG heartbeat signal, and the BCG dataset comprises at least a BCG respiratory signal and a BCG heartbeat signal; pre-training a sleep monitoring initial model based on the PSG dataset to obtain a sleep monitoring base model; and fine-tuning the sleep monitoring base model based on the BCG dataset to obtain the sleep monitoring model.

2. The method of claim 1, wherein the sleep monitoring model is configured to perform at least a sleep staging task and / or a sleep apnea estimation task. The pre-training of the sleep monitoring initial model based on the PSG dataset comprises:

3. The method of claim 1, wherein, training using the PSG respiratory signal and the PSG heartbeat signal as a training set. The fine-tuning of the sleep monitoring base model based on the BCG dataset comprises:

4. The method of claim 1, wherein, fine-tuning the sleep monitoring base model using the BCG respiratory signal and the BCG heartbeat signal, and additional bandpass signals, wherein the additional bandpass signals comprise a bandpass filtered BCG respiratory signal and a bandpass filtered BCG heartbeat signal. The additional bandpass signals further comprise a bandpass filtered BCG snoring signal.

5. The method of claim 4, wherein, The sleep monitoring model comprises a CNN feature extraction module, an adaptive feature fusion module, and a Roformer encoding module, and wherein:

6. The method of claim 5, wherein, the CNN feature extraction module is configured to perform feature extraction on the BCG respiratory signal and the BCG heartbeat signal, and the bandpass filtered BCG respiratory signal, the bandpass filtered BCG heartbeat signal, and the bandpass filtered BCG snoring signal, respectively, to obtain enhanced features of the BCG respiratory signal and the BCG heartbeat signal, and bandpass features of the bandpass filtered BCG respiratory signal, the bandpass filtered BCG heartbeat signal, and the bandpass filtered BCG snoring signal; the adaptive feature fusion module is configured to perform feature fusion on the enhanced features and the bandpass features to obtain fused features; and the Roformer encoding module is configured to process the fused features to perform sleep staging or sleep apnea estimation. The sleep monitoring model further comprises an Adam optimizer configured to dynamically adjust a learning rate for the sleep staging task and the sleep apnea estimation task.

7. The method of claim 6, wherein, The sleep monitoring model is configured to output a sleep apnea severity, including mild, moderate, and severe, by estimating an AHI value.

8. The method of claim 2, wherein, The PSG heartbeat signal comprises an RRI heartbeat sequence, and the BCG heartbeat signal comprises a JJI heartbeat sequence.

9. The method of claim 1, wherein, ​ 10. The method of claim 1, further comprising performing signal quality assessment on the PSG dataset and the BCG dataset, and selecting the PSG respiratory signal and the PSG heartbeat signal for pre-training and the BCG respiratory signal and the BCG heartbeat signal for fine-tuning based on the assessment results.

11. The method of claim 10, wherein, The signal quality assessment comprises: calculating a confidence score, which is the ratio between the sum of the energy of the PSG or BCG respiratory signal and the PSG or BCG heartbeat signal in each time segment and the total energy in the time segment; and calculating an average confidence score over all time segments, wherein, in the case that the average confidence score is greater than a threshold, the PSG or BCG respiratory signal and the PSG or BCG heartbeat signal are retained for pre-training or fine-tuning of the sleep monitoring base model, and wherein, in the case that the average confidence score is less than or equal to the threshold, the PSG or BCG respiratory signal and the PSG or BCG heartbeat signal are discarded.

12. The method of claim 11, wherein, The threshold is 0.

6.

13. The method of any one of claims 1 to 12, wherein, The PSG dataset is obtained from a plurality of sources, including one or more of SHHS, MESA, MROS, CHAT, WSC, PATS, NUMOM2B, and HSP.

14. A sleep monitoring method, the method comprising: acquiring a BCG respiratory signal and a BCG heartbeat signal of a user; and analyzing the BCG respiratory signal and the BCG heartbeat signal using a sleep monitoring model obtained according to any one of claims 1 to 13 to obtain sleep staging results and AHI severity for the user.

15. An apparatus for training a sleep monitoring model based on transfer learning, the apparatus comprising: means for obtaining a polysomnography (PSG) dataset and a ballistocardiogram (BCG) dataset, wherein the PSG dataset comprises at least a PSG respiratory signal and a PSG heartbeat signal, and the BCG dataset comprises at least a BCG respiratory signal and a BCG heartbeat signal; means for pre-training a sleep monitoring initial model based on the PSG dataset to obtain a sleep monitoring base model; means for fine-tuning the sleep monitoring base model based on the BCG dataset to obtain the sleep monitoring model.

16. A sleep monitoring apparatus, the apparatus comprising: means for acquiring a BCG respiratory signal and a BCG heartbeat signal of a user; and means for analyzing the BCG respiratory signal and the BCG heartbeat signal using a sleep monitoring model obtained according to any one of claims 1 to 13 to obtain sleep staging results and AHI severity for the user.

17. An electronic device comprising the sleep monitoring apparatus of claim 16.

18. A computer device comprising a memory and a processor, wherein, The memory has stored therein computer readable instructions which, when executed by the processor, implement the method of any one of claims 1 to 14.

18. A computer program product comprising a computer readable medium having stored therein the computer readable instructions of claim 17.

19. A computer-readable storage medium having stored thereon computer-readable instructions which, when executed by a processor, implement the method of any one of claims 1 to 14.

Images