Multimodal sleep evaluation system based on teacher-student model

Abstract

The invention discloses a multi-mode sleep evaluation system based on a teacher-student model. The system comprises a signal acquisition module used for synchronously acquiring multi-modal physiological signals of a to-be-evaluated object in a sleep period; the preprocessing module is used for preprocessing the multi-modal physiological signals to remove abnormal values and obtain standardized multi-modal physiological signal data so as to meet the input requirements of a teacher-student model; the model construction module is used for constructing a teacher-student knowledge distillation framework which comprises a teacher model and a student model; and the evaluation module is used for deploying the trained student model in a reasoning stage, receiving the preprocessed multi-modal physiological signal and synchronously outputting a sleep staging result and an apnea hypopnea index evaluation result. According to the method, the problem of insufficient sleep data annotation can be effectively relieved, the evaluation precision and reliability of light-contact equipment are improved, and the method has wide clinical auxiliary diagnosis and long-term health management application value.

Description

Technical Field

[0001] This invention belongs to the field of intelligent health monitoring technology, and in particular to a multimodal sleep assessment system based on a teacher-student model. Background Technology

[0002] Sleep is one of the most basic physiological needs of the human body, and its quality is closely related to individual health. According to the World Health Organization, about one-third of the global population suffers from varying degrees of sleep disorders, such as insomnia and sleep apnea. These problems not only affect quality of life but also significantly increase the risk of cardiovascular disease, metabolic syndrome, and cognitive impairment. Therefore, conducting accurate and convenient sleep quality assessments is of great significance for the early diagnosis and health intervention of sleep disorders.

[0003] Currently, the gold standard for clinical diagnosis of sleep disorders is polysomnography (PSG). PSG records more than ten physiological parameters, such as electroencephalography (EEG), electrooculography (EOG), electromyography (EMG), electrocardiography (ECG), respiratory effort, and blood oxygen saturation, by attaching multiple electrodes and sensors to the subject's body. With the assistance of a professional physician, sleep stages and respiratory events are manually interpreted according to standards. Although PSG results are accurate, it has obvious limitations: (1) It is highly invasive and uncomfortable: a large number of wires and sensors can cause discomfort to the subject and may interfere with normal sleep, producing a "first night effect" and affecting the authenticity of the results; (2) It must be carried out in a professional sleep laboratory, the equipment is expensive, it requires a specialist to operate, the cost is high, and the accessibility is poor; (3) Data analysis depends on manual labor, which is time-consuming, laborious, subjective, and inefficient.

[0004] To overcome the above shortcomings, home sleep monitoring solutions based on portable devices such as smart bracelets and rings have emerged in recent years. These devices are usually based on photoplethysmography (PPG) and accelerometer signals, and use algorithms to estimate sleep stages and blood oxygen levels. However, such solutions have the following problems: (1) Limited accuracy: PPG signals are easily affected by motion artifacts and cannot effectively distinguish sleep stages, especially the deep stages of rapid eye movement (REM) sleep and non-rapid eye movement (NREM) sleep; (2) Limited functionality: Most consumer-grade devices are difficult to accurately detect events such as sleep apnea, or the reliability of the apnea-hypopnea index (AHI) is insufficient, which cannot meet the needs of clinical auxiliary diagnosis.

[0005] In recent years, bio-radar technology has been explored for monitoring breathing and heart rate during sleep due to its non-contact and highly sensitive characteristics. However, when used alone, it still faces challenges in terms of the comprehensiveness and stability of vital sign measurements.

[0006] In summary, a key contradiction exists in the current field of sleep monitoring: PSG offers high accuracy but is not readily available, while portable devices are highly accessible but lack sufficient accuracy. Therefore, there is an urgent need for a new sleep assessment solution that can balance high accuracy with home convenience. Summary of the Invention

[0007] The purpose of this invention is to address the shortcomings and deficiencies of existing technologies by providing a multimodal sleep assessment system based on a teacher-student model. This system aims to solve the problems of invasiveness, high cost, and low accessibility of PSG monitoring, while overcoming the insufficient accuracy of single portable devices. It aims to provide a convenient, comfortable, and clinically reference-level accuracy sleep quality assessment solution for home environments, effectively solving the industry challenges of insufficient accuracy in sleep monitoring methods based on light-contact devices and the difficulty in acquiring large-scale PSG labeled data.

[0008] The technical solution for achieving the objective of this invention is: a multimodal sleep assessment system based on a teacher-student model, the system comprising:

[0009] The signal acquisition module is used to synchronously acquire multimodal physiological signals of the subject under evaluation during sleep.

[0010] The preprocessing module is used to preprocess multimodal physiological signals to remove outliers and obtain standardized multimodal physiological signal data to meet the input requirements of the teacher-student model.

[0011] The model building module is used to build a knowledge distillation framework for teachers and students, including teacher models and student models;

[0012] The evaluation module is used to deploy the trained student model during the inference phase, receive preprocessed multimodal physiological signals, and simultaneously output sleep staging results and apnea-hypopnea index evaluation results.

[0013] Furthermore, the signal acquisition module includes:

[0014] The first unit is used to collect physiological signals of the subject to be evaluated through a polysomnography device, including at least electrocardiogram (ECG) signals, respiratory signals, and blood oxygenation signals.

[0015] The second unit is used to collect vital signs information, including at least respiratory signals, via bio-radar; and simultaneously, it is used to collect physiological signals of the subject to be evaluated, including at least pulse signals and blood oxygen signals, via a ring-type wearable device.

[0016] Furthermore, the bio-radar is placed directly above the object to be evaluated, aimed at the chest and abdomen of the object, and the monitoring distance is 2.5~3 m.

[0017] Furthermore, the bio-radar acquires data from a frequency-modulated continuous wave radar.

[0018] Furthermore, the specific process of the bio-radar acquiring respiratory signals includes:

[0019] The received echo signal containing vital signs information is processed to obtain vital signs signals at different distances:

[0020]

[0021] in, Let be the set of all range-gate vital sign signals at time t, and n be the number of range gates. The vital sign signal corresponding to the p-th distance gate;

[0022] The maximum ratio combination-principal component analysis method was used to fuse vital sign signals from multiple distance gates. The specific process included:

[0023] (1) Calculate the signal-to-noise ratio of vital signs signals at each distance gate:

[0024]

[0025] in, For the p-th distance gate signal The signal-to-noise ratio, For the p-th distance gate signal Signal power in the 0.1Hz ~ 0.6Hz frequency band, For the p-th distance gate signal Signal power in the frequency band above 0.6 Hz;

[0026] (2) Using principal component analysis, the first principal component of the vital signs signal under each distance gate is calculated, and the final gain value of each vital signs signal under each distance gate is obtained based on the sign of the first principal component and the signal-to-noise ratio:

[0027]

[0028] in, The first principal component of the p-th distance gate signal, Let be the final gain of the p-th distance gate signal;

[0029] (3) Using the maximum ratio combination method, based on the final gain of all range gate signals, the vital signs signals under all range gates are fused into a single vital signs signal by calculating a weighted average. :

[0030]

[0031] Furthermore, the preprocessing module includes:

[0032] The time alignment unit is used to: based on the shared cardiac cycle characteristics of the pulse signal collected by the ring-type wearable device and the vital sign information collected by the bio-radar, determine the time delay of the bio-radar signal transmission relative to the signal transmission of the ring-type wearable device by calculating the position of the maximum value of the cross-correlation function, and perform corresponding time shift correction on the bio-radar signal transmission.

[0033] The signal length standardization unit is used to uniformly process the length of the multimodal physiological signal to a preset duration; to truncate signal segments with a duration exceeding the preset duration and to fill signal segments with insufficient duration to form signal data with a uniform duration.

[0034] The filtering and data cleaning unit is used to filter the signal after it has been processed by the signal length normalization unit to remove high-frequency noise and baseline drift, and to clean or mark invalid or artifact-prone signal segments based on the signal quality detection results.

[0035] The resampling unit is used to unify the sampling rate of various signals after they have been processed by the filtering and data cleaning units to the target frequency;

[0036] The data standardization unit is used to normalize the amplitude of the signal processed by the resampling unit, so that its numerical range meets the specifications of the subsequent model input, and generates the final standardized multimodal physiological signal data.

[0037] Furthermore, the aforementioned teacher-student knowledge distillation framework specifically includes:

[0038] The teacher model's core is a cross-modal context encoder. This encoder adopts a symmetrical dual-module structure and uses a cross-attention mechanism to enhance the connection between ECG features and respiratory features, as well as the connection between respiratory features and ECG features, to deeply integrate the correlation information between different physiological modalities. The teacher model takes standardized data generated by the preprocessing module from respiratory signals, ECG signals, and blood oxygen signals collected by polysomnography as input, and outputs high-precision sleep staging results and apnea-hypopnea index estimates, i.e., AHI estimates.

[0039] The student model has the same network structure as the teacher model. The student model takes standardized data generated by the preprocessing module from the respiratory signal, pulse signal and blood oxygen signal from the bio-radar and the ring-type wearable device as input, and outputs high-precision sleep stage results and apnea-hypopnea index estimates, i.e., AHI estimates.

[0040] The knowledge transfer module connects the teacher model and the student model. During the training phase, it drives the student model to learn from the teacher model at both the feature and output levels by jointly optimizing multiple loss functions. The loss functions include at least: Euclidean distance loss for constraining the similarity of intermediate features, maximum mean difference loss for aligning feature distributions, and Kullback-Leibler divergence loss for aligning output probability distributions.

[0041] Furthermore, the teacher model includes:

[0042] Three convolutional neural networks are used to extract the respiratory, electrocardiographic, and blood oxygenation features of the respiratory, ECG, and blood oxygenation signals collected by the polysomnography device, respectively.

[0043] A cross-modal context encoder for enhancing and fusing respiratory and electrocardiogram features;

[0044] A sleep analysis unit is used to perform sleep staging based on features output by a cross-modal context encoder; the unit includes a first dilated convolutional block, a first multi-head attention block, a second dilated convolutional block, a second multi-head attention block, and a classifier connected in sequence.

[0045] The splicing unit is used to splice the features output by the cross-modal context encoder and the blood oxygen feature;

[0046] The AHI estimation unit is used to predict AHI estimates based on the features output by the splicing unit. The unit includes a convolutional block, an AHI prediction block, and a fully connected layer connected in sequence. The AHI prediction block is used to regress and calculate the mean apnea-hypopnea index.

[0047] Furthermore, the AHI prediction block includes four stacked dilated convolutional blocks and a subsequent fully connected network. Each dilated convolutional block includes three dilated convolutional layers, one batch normalization layer, and one max pooling layer. The dilation rates of the dilated convolutional blocks are set to 1, 2, 3, and 4, respectively, and the number of convolutional kernels are 64, 64, 32, and 16, respectively, with a uniform kernel size of 9. The window size and stride of the max pooling layer are both 2. The AHI prediction block gradually extracts multi-scale features and compresses the feature dimensions through the dilated convolutional blocks. Its output is converted into a one-dimensional feature vector by a flattening layer and then fed into the fully connected network for information integration. Subsequently, it is gradually reduced in dimensionality through three consecutive fully connected layers and a batch normalization layer, and finally mapped to a single AHI estimate.

[0048] Furthermore, the evaluation module includes:

[0049] The sleep staging unit is used to automatically segment overnight sleep data based on deep features extracted from the student model, and output classification results including wakefulness, REM sleep, light sleep and deep sleep.

[0050] The AHI estimation unit is used to calculate the average apnea-hypopnea index during the entire night's sleep based on the deep features extracted from the student model, and outputs an AHI estimate.

[0051] Compared with the prior art, the significant advantages of this invention are:

[0052] (1) The sleep assessment system proposed in this invention transfers the clinical-grade analysis capabilities of PSG to a convenient model based on radar and rings through knowledge distillation technology, achieving comfortable home-based monitoring while ensuring high accuracy.

[0053] (2) It creatively combines the complementary characteristics of two modes, namely bio-radar (non-contact and strong anti-interference) and ring (direct and accurate blood oxygen measurement), and constructs a more comprehensive and reliable vital sign information system than a single sensor through signal-level alignment and model-level fusion, which significantly enhances the stability and anti-interference ability of the system in a real home environment.

[0054] (3) From multimodal signal preprocessing and feature extraction to the final sleep staging and AHI estimation, the entire process is fully automated, avoiding the time-consuming and laborious manual interpretation in traditional PSG analysis, greatly improving the efficiency of sleep assessment, and providing a feasible technical path for long-term sleep health screening and monitoring of large-scale populations.

[0055] The present invention will now be described in further detail with reference to the accompanying drawings. Attached Figure Description

[0056] Figure 1 This is a flowchart illustrating the overall architecture of the multimodal sleep assessment system based on a teacher-student model, as described in this invention.

[0057] Figure 2 This is a detailed structural diagram of the teacher model in one embodiment.

[0058] Figure 3 This is a detailed diagram of a cross-modal attention fusion module in one embodiment.

[0059] Figure 4 This is a detailed schematic diagram of a teacher-student knowledge distillation framework in one embodiment.

[0060] Figure 5 Here is an example diagram of multi-task output (sleep staging + AHI estimation) in one embodiment, where Figure 5 (a) in the figure is a whole night sleep analysis chart. Figure 5 (b) in the figure is a two-dimensional scatter plot of AHI. Figure 5 (c) in the figure is the Bland-Altman plot of AHI. Figure 5(d) in the figure represents the confusion matrix for the severity of sleep-disordered breathing. Detailed Implementation

[0061] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0062] It should be noted that if the embodiments of the present invention involve directional indicators (such as up, down, left, right, front, back, etc.), the directional indicators are only used to explain the relative positional relationship and movement of the components in a certain specific posture (as shown in the figure). If the specific posture changes, the directional indicators will also change accordingly.

[0063] Furthermore, if the embodiments of this invention involve descriptions such as "first" or "second," these descriptions are for descriptive purposes only and should not be construed as indicating or implying their relative importance or implicitly specifying the number of technical features indicated. Therefore, a feature defined with "first" or "second" may explicitly or implicitly include at least one of those features. Additionally, the technical solutions of the various embodiments can be combined with each other, but this must be based on the ability of those skilled in the art to implement them. If the combination of technical solutions is contradictory or impossible to implement, it should be considered that such a combination of technical solutions does not exist and is not within the scope of protection claimed by this invention.

[0064] In one embodiment, a multimodal sleep assessment system based on a teacher-student model is provided, the system comprising:

[0065] The signal acquisition module is used to synchronously acquire multimodal physiological signals of the subject under evaluation during sleep.

[0066] The preprocessing module is used to preprocess multimodal physiological signals to remove outliers and obtain standardized multimodal physiological signal data to meet the input requirements of the teacher-student model.

[0067] The model building module is used to build a knowledge distillation framework for teachers and students, including teacher models and student models;

[0068] The evaluation module is used to deploy the trained student model during the inference phase, receive preprocessed multimodal physiological signals, and simultaneously output sleep staging results and apnea-hypopnea index evaluation results.

[0069] Furthermore, in one embodiment, the signal acquisition module includes:

[0070] The first unit is used to collect physiological signals of the subject to be evaluated through a polysomnography device, including at least electrocardiogram (ECG) signals, respiratory signals, and blood oxygenation signals.

[0071] Specifically, in accordance with the standard polysomnography monitoring specifications, electrodes and sensors are attached to specific parts of the user's body to collect raw physiological signals; the physiological signals include at least lead electrocardiogram (ECG) signals, respiratory effort waveforms, and blood oxygen saturation (SpO2) signals; wherein, the ECG signal is collected through chest lead electrodes, the respiratory effort waveform is collected through a respiratory plethysmography sensor wrapped around the chest and abdomen, and the SpO2 signal is collected through a fingertip pulse oximeter.

[0072] The second unit is used to collect vital signs information, including at least respiratory signals, via bio-radar; and simultaneously, it is used to collect physiological signals of the subject to be evaluated, including at least pulse signals and blood oxygen signals, via a ring-type wearable device.

[0073] Preferably, the bio-radar is placed directly above the object to be evaluated, aimed at the chest and abdomen of the object, and the monitoring distance is 2.5~3 m.

[0074] Preferably, the bio-radar acquires data from a frequency-modulated continuous wave radar.

[0075] Preferably, the specific process of the bio-radar acquiring respiratory signals includes:

[0076] The received echo signal containing vital signs information is processed to obtain vital signs signals at different distances:

[0077]

[0078] in, Let be the set of all range-gate vital sign signals at time t, and n be the number of range gates. The vital sign signal corresponding to the p-th distance gate;

[0079] The maximum ratio combination-principal component analysis method was used to fuse vital sign signals from multiple distance gates. The specific process included:

[0080] (1) Calculate the signal-to-noise ratio (SNR) of vital signs at each distance point. The SNR is calculated as the ratio of the signal power within the normal respiratory rate range (0.1 Hz ~ 0.6 Hz) to the signal power above the normal respiratory rate.

[0081]

[0082] in, For the p-th distance gate signal The signal-to-noise ratio, For the p-th distance gate signal Signal power in the 0.1Hz ~ 0.6Hz frequency band, For the p-th distance gate signal Signal power in the frequency band above 0.6 Hz;

[0083] (2) Using principal component analysis, the first principal component of the vital signs signal under each distance gate is calculated, and the final gain value of each vital signs signal under each distance gate is obtained based on the sign of the first principal component and the signal-to-noise ratio:

[0084]

[0085] in, The first principal component of the p-th distance gate signal, Let be the final gain of the p-th distance gate signal;

[0086] (3) Using the maximum ratio combination method, based on the final gain of all range gate signals, the vital signs signals under all range gates are fused into a single vital signs signal by calculating a weighted average. :

[0087]

[0088] Furthermore, in one embodiment, the preprocessing module includes:

[0089] The time alignment unit is used to: based on the shared cardiac cycle characteristics of the pulse signal collected by the ring-type wearable device and the vital sign information collected by the bio-radar, determine the time delay of the bio-radar signal transmission relative to the signal transmission of the ring-type wearable device by calculating the position of the maximum value of the cross-correlation function, and perform corresponding time shift correction on the bio-radar signal transmission.

[0090] A signal length standardization unit is used to uniformly process the length of the multimodal physiological signal to a preset duration; to truncate signal segments with a duration exceeding the preset duration and to fill signal segments with a duration insufficient to form signal data with a uniform duration; preferably, the preset duration is 10 hours;

[0091] The filtering and data cleaning unit is used to filter the signal after it has been processed by the signal length normalization unit to remove high-frequency noise and baseline drift, and to clean or mark invalid or artifact-prone signal segments based on the signal quality detection results.

[0092] The resampling unit is used to unify the sampling rate of various signals after they have been processed by the filtering and data cleaning units to the target frequency;

[0093] The data standardization unit is used to normalize the amplitude of the signal processed by the resampling unit, so that its numerical range meets the specifications of the subsequent model input, and generates the final standardized multimodal physiological signal data.

[0094] Furthermore, in one embodiment, the teacher-student knowledge distillation framework specifically includes:

[0095] Teacher models, such as Figure 2 As shown, the core is a cross-modal context encoder, such as... Figure 3 As shown, the encoder adopts a symmetrical dual-module structure, and enhances the ECG features to respiratory features and the respiratory features to ECG features through a cross-attention mechanism, so as to deeply integrate the correlation information between different physiological modalities. The teacher model takes the standardized data generated by the preprocessing module of the respiratory signal, ECG signal and blood oxygen signal collected by the polysomnography device as input, and outputs high-precision sleep staging results and apnea-hypopnea index estimates, i.e., AHI estimates.

[0096] The student model has the same network structure as the teacher model to ensure consistency in cross-modal feature transfer. The student model takes standardized data generated by the preprocessing module from the respiratory signal, pulse signal and blood oxygen signal from the bio-radar and the ring-type wearable device as input, and outputs high-precision sleep staging results and apnea-hypopnea index estimates, i.e., AHI estimates.

[0097] The knowledge transfer module, through joint optimization of multiple loss functions, drives the student model to mimic the teacher model. Its framework is as follows: Figure 4 As shown. Connecting the teacher model and the student model, during the training phase, the student model is driven to learn from the teacher model at both the feature and output levels by jointly optimizing multiple loss functions; the loss functions include at least: Euclidean distance loss for constraining the similarity of intermediate features, maximum mean difference loss for aligning feature distributions, and Kullback-Leibler (KL) divergence loss for aligning output probability distributions.

[0098] Preferably, the teacher model includes:

[0099] Three convolutional neural networks are used to extract the respiratory, electrocardiographic, and blood oxygenation features of the respiratory, ECG, and blood oxygenation signals collected by the polysomnography device, respectively.

[0100] A cross-modal context encoder for enhancing and fusing respiratory and electrocardiogram features;

[0101] A sleep analysis unit is used to perform sleep staging based on features output by a cross-modal context encoder; the unit includes a first dilated convolutional block, a first multi-head attention block, a second dilated convolutional block, a second multi-head attention block, and a classifier connected in sequence.

[0102] The splicing unit is used to splice the features output by the cross-modal context encoder and the blood oxygen feature;

[0103] The AHI estimation unit is used to predict AHI estimates based on the features output by the splicing unit. The unit includes a convolutional block, an AHI prediction block, and a fully connected layer connected in sequence. The AHI prediction block is used to regress and calculate the mean apnea-hypopnea index.

[0104] Preferably, the AHI prediction block includes four stacked dilated convolutional blocks and a subsequent fully connected network; each dilated convolutional block includes three dilated convolutional layers, one batch normalization layer, and one max pooling layer; the dilation rates of the dilated convolutional blocks are set to 1, 2, 3, and 4 respectively, the number of convolutional kernels are 64, 64, 32, and 16 respectively, and the kernel size is uniformly 9; the window size and stride of the max pooling layer are both 2; the AHI prediction block gradually extracts multi-scale features and compresses the feature dimensions through the dilated convolutional blocks, and its output is converted into a one-dimensional feature vector by a flattening layer, and then fed into the fully connected network for information integration. Subsequently, it gradually reduces the dimensionality through three consecutive fully connected layers and a batch normalization layer, and finally maps to a single AHI estimate.

[0105] Furthermore, in one embodiment, the evaluation module includes:

[0106] The sleep staging unit is used to automatically segment overnight sleep data based on deep features extracted by the student model, and output classification results including wakefulness, rapid eye movement sleep (REM), light sleep, and deep sleep.

[0107] The AHI estimation unit is used to calculate the average apnea-hypopnea index during the entire night's sleep based on the deep features extracted from the student model, and outputs an AHI estimate.

[0108] As a specific example, the invention is illustrated in one embodiment.

[0109] For a specific assessment subject, the overnight sleep analysis chart is as follows: Figure 5 As shown in (a) of the figure, it can be seen that the model's prediction results are generally consistent with the actual sleep labels, and obvious trends in sleep cycle changes can be identified.

[0110] For the AHI estimation task, this embodiment also analyzed the results. The MAE between the actual AHI value and the estimated AHI value was 5.55, the ICC was 0.86, and the Pearson correlation coefficient was 0.89. Figure 5 The results of the AHI estimation task are further presented. From Figure 5 As can be seen in (b) of the data, there is good consistency between the actual AHI value and the estimated AHI value. From... Figure 5 As can be seen in (c), the average difference between the actual AHI value and the estimated AHI value is 3.03, which is close to 0, and most values ​​fall within the 95% confidence interval, which also indicates that there is good consistency between the two. Figure 5 (d) in the figure represents the confusion matrix for classifying the severity of sleep-disordered breathing based on the AHI score. The confusion matrix shows an accuracy of 60.03%, an F1 score of 0.612, a Kappa score of 0.451, a recall of 0.623, and a precision of 0.647 for severity levels of normal, mild, moderate, and severe sleep-disordered breathing. The recall rates for these levels are 0.812, 0.526, 0.605, and 0.550, respectively. The results indicate that the AHI estimation task can also achieve good results.

[0111] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely illustrative of the principles of the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the claimed invention.

Claims

1. A multimodal sleep assessment system based on a teacher-student model, characterized in that, The system includes: The signal acquisition module is used to synchronously acquire multimodal physiological signals of the subject under evaluation during sleep. The preprocessing module is used to preprocess multimodal physiological signals to remove outliers and obtain standardized multimodal physiological signal data to meet the input requirements of the teacher-student model. The model building module is used to build a knowledge distillation framework for teachers and students, including teacher models and student models; The evaluation module is used to deploy the trained student model during the inference phase, receive preprocessed multimodal physiological signals, and simultaneously output sleep staging results and apnea-hypopnea index evaluation results.

2. The multimodal sleep assessment system based on a teacher-student model according to claim 1, characterized in that, The signal acquisition module includes: The first unit is used to collect physiological signals of the subject to be evaluated through a polysomnography device, including at least electrocardiogram (ECG) signals, respiratory signals, and blood oxygenation signals. The second unit is used to collect vital signs information, including at least respiratory signals, via bio-radar; and simultaneously, it is used to collect physiological signals of the subject to be evaluated, including at least pulse signals and blood oxygen signals, via a ring-type wearable device.

3. The multimodal sleep assessment system based on a teacher-student model according to claim 2, characterized in that, The bio-radar is positioned directly above the object to be evaluated, aimed at the chest and abdomen, and has a monitoring distance of 2.5 to 3 meters.

4. The multimodal sleep assessment system based on a teacher-student model according to claim 2, characterized in that, The bio-radar acquires data from a frequency-modulated continuous wave radar.

5. The multimodal sleep assessment system based on a teacher-student model according to claim 2, characterized in that, The specific process of the bio-radar acquiring respiratory signals includes: The received echo signal containing vital signs information is processed to obtain vital signs signals at different distances: ； in, Let be the set of all range-gate vital sign signals at time t, and n be the number of range gates. The vital sign signal corresponding to the p-th distance gate; The maximum ratio combination-principal component analysis method was used to fuse vital sign signals from multiple distance gates. The specific process included: (1) Calculate the signal-to-noise ratio of vital signs signals at each distance gate: ； in, For the p-th distance gate signal The signal-to-noise ratio, For the p-th distance gate signal Signal power in the 0.1 Hz ~ 0.6 Hz frequency band, For the p-th distance gate signal Signal power in the frequency band above 0.6 Hz; (2) Using principal component analysis, the first principal component of the vital signs signal under each distance gate is calculated, and the final gain value of each vital signs signal under each distance gate is obtained based on the sign of the first principal component and the signal-to-noise ratio: ； in, The first principal component of the p-th distance gate signal is... Let be the final gain of the p-th distance gate signal; (3) Using the maximum ratio combination method, based on the final gain of all range gate signals, the vital signs signals under all range gates are fused into a single vital signs signal by calculating a weighted average. : 。 6. The multimodal sleep assessment system based on a teacher-student model according to claim 2, characterized in that, The preprocessing module includes: The time alignment unit is used to: based on the shared cardiac cycle characteristics of the pulse signal collected by the ring-type wearable device and the vital sign information collected by the bio-radar, determine the time delay of the bio-radar signal transmission relative to the signal transmission of the ring-type wearable device by calculating the position of the maximum value of the cross-correlation function, and perform corresponding time shift correction on the bio-radar signal transmission. The signal length standardization unit is used to uniformly process the length of the multimodal physiological signal to a preset duration; to truncate signal segments with a duration exceeding the preset duration and to fill signal segments with insufficient duration to form signal data with a uniform duration. The filtering and data cleaning unit is used to filter the signal after it has been processed by the signal length normalization unit to remove high-frequency noise and baseline drift, and to clean or mark invalid or artifact-prone signal segments based on the signal quality detection results. The resampling unit is used to unify the sampling rate of various signals after they have been processed by the filtering and data cleaning units to the target frequency; The data standardization unit is used to normalize the amplitude of the signal processed by the resampling unit, so that its numerical range meets the specifications of the subsequent model input, and generates the final standardized multimodal physiological signal data.

7. The multimodal sleep assessment system based on a teacher-student model according to claim 2, characterized in that, The aforementioned knowledge distillation framework for teachers and students specifically includes: The teacher model's core is a cross-modal context encoder. This encoder adopts a symmetrical dual-module structure and uses a cross-attention mechanism to enhance the connection between ECG features and respiratory features, as well as the connection between respiratory features and ECG features, to deeply integrate the correlation information between different physiological modalities. The teacher model takes standardized data generated by the preprocessing module from respiratory signals, ECG signals, and blood oxygen signals collected by polysomnography as input, and outputs high-precision sleep staging results and apnea-hypopnea index estimates, i.e., AHI estimates. The student model has the same network structure as the teacher model. The student model takes standardized data generated by the preprocessing module from the respiratory signal, pulse signal and blood oxygen signal from the bio-radar and the ring-type wearable device as input, and outputs high-precision sleep stage results and apnea-hypopnea index estimates, i.e., AHI estimates. The knowledge transfer module connects the teacher model and the student model. During the training phase, it drives the student model to learn from the teacher model at both the feature and output levels by jointly optimizing multiple loss functions. The loss functions include at least: Euclidean distance loss for constraining the similarity of intermediate features, maximum mean difference loss for aligning feature distributions, and Kullback-Leibler divergence loss for aligning output probability distributions.

8. The multimodal sleep assessment system based on a teacher-student model according to claim 7, characterized in that, The teacher model includes: Three convolutional neural networks are used to extract the respiratory, electrocardiographic, and blood oxygenation features of the respiratory, ECG, and blood oxygenation signals collected by the polysomnography device, respectively. A cross-modal context encoder for enhancing and fusing respiratory and electrocardiogram features; A sleep analysis unit is used to perform sleep staging based on features output by a cross-modal context encoder; the unit includes a first dilated convolutional block, a first multi-head attention block, a second dilated convolutional block, a second multi-head attention block, and a classifier connected in sequence. The splicing unit is used to splice the features output by the cross-modal context encoder and the blood oxygen feature; The AHI estimation unit is used to predict AHI estimates based on the features output by the splicing unit. The unit includes a convolutional block, an AHI prediction block, and a fully connected layer connected in sequence. The AHI prediction block is used to regress and calculate the mean apnea-hypopnea index.

9. The multimodal sleep assessment system based on a teacher-student model according to claim 8, characterized in that, The AHI prediction block comprises four stacked dilated convolutional blocks and a subsequent fully connected network. Each dilated convolutional block includes three dilated convolutional layers, one batch normalization layer, and one max pooling layer. The dilation rates of the dilated convolutional blocks are set to 1, 2, 3, and 4, respectively, with 64, 64, 32, and 16 convolutional kernels, and a uniform kernel size of 9. The window size and stride of the max pooling layer are both 2. This AHI prediction block extracts multi-scale features and compresses the feature dimensions step by step through the dilated convolutional blocks. Its output is converted into a one-dimensional feature vector by a flattening layer and then fed into the fully connected network for information integration. Subsequently, it is gradually reduced in dimensionality through three consecutive fully connected layers and a batch normalization layer, ultimately mapping to a single AHI estimate.

10. The multimodal sleep assessment system based on a teacher-student model according to claim 1, characterized in that, The evaluation module includes: The sleep staging unit is used to automatically segment overnight sleep data based on deep features extracted from the student model, and output classification results including wakefulness, REM sleep, light sleep and deep sleep. The AHI estimation unit is used to calculate the average apnea-hypopnea index during the entire night's sleep based on the deep features extracted from the student model, and outputs an AHI estimate.

Images